Upload folder using huggingface_hub

.gitignore

@@ -0,0 +1,160 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

Bert_VITS2_Guide.ipynb

@@ -0,0 +1,323 @@
+{
+  "nbformat": 4,
+  "nbformat_minor": 0,
+  "metadata": {
+    "colab": {
+      "provenance": [],
+      "authorship_tag": "ABX9TyP+PQNqeOew0Ap+hzVH7i8r",
+      "include_colab_link": true
+    },
+    "kernelspec": {
+      "name": "python3",
+      "display_name": "Python 3"
+    },
+    "language_info": {
+      "name": "python"
+    }
+  },
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "view-in-github",
+        "colab_type": "text"
+      },
+      "source": [
+        "<a href=\"https://colab.research.google.com/github/KevinWang676/Bert-VITS2-quick-start/blob/main/Bert_VITS2_Guide.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "### 0. 如果使用AutoDL，请运行下载packages的加速代码："
+      ],
+      "metadata": {
+        "id": "CGg4SV4ObQaT"
+      }
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "MgfAJzoHbK2-"
+      },
+      "outputs": [],
+      "source": [
+        "!source /etc/network_turbo\n",
+        "!python -m pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "### 1. 数据集重采样和标注"
+      ],
+      "metadata": {
+        "id": "sloMn00-bgxY"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "import subprocess\n",
+        "import random\n",
+        "import os\n",
+        "from pathlib import Path\n",
+        "import librosa\n",
+        "from scipy.io import wavfile\n",
+        "import numpy as np\n",
+        "import torch\n",
+        "import csv\n",
+        "import whisper\n",
+        "\n",
+        "a=\"linghua\" # 请在这里修改说话人的名字，目前只支持中文语音\n",
+        "\n",
+        "def split_long_audio(model, filepaths, save_dir=\"data_dir\", out_sr=44100):\n",
+        "    if isinstance(filepaths, str):\n",
+        "        filepaths = [filepaths]\n",
+        "\n",
+        "    for file_idx, filepath in enumerate(filepaths):\n",
+        "\n",
+        "        save_path = Path(save_dir)\n",
+        "        save_path.mkdir(exist_ok=True, parents=True)\n",
+        "\n",
+        "        print(f\"Transcribing file {file_idx}: '{filepath}' to segments...\")\n",
+        "        result = model.transcribe(filepath, word_timestamps=True, task=\"transcribe\", beam_size=5, best_of=5)\n",
+        "        segments = result['segments']\n",
+        "\n",
+        "        wav, sr = librosa.load(filepath, sr=None, offset=0, duration=None, mono=True)\n",
+        "        wav, _ = librosa.effects.trim(wav, top_db=20)\n",
+        "        peak = np.abs(wav).max()\n",
+        "        if peak > 1.0:\n",
+        "            wav = 0.98 * wav / peak\n",
+        "        wav2 = librosa.resample(wav, orig_sr=sr, target_sr=out_sr)\n",
+        "        wav2 /= max(wav2.max(), -wav2.min())\n",
+        "\n",
+        "        for i, seg in enumerate(segments):\n",
+        "            start_time = seg['start']\n",
+        "            end_time = seg['end']\n",
+        "            wav_seg = wav2[int(start_time * out_sr):int(end_time * out_sr)]\n",
+        "            wav_seg_name = f\"{a}_{i}.wav\" # 在上方可修改名字\n",
+        "            out_fpath = save_path / wav_seg_name\n",
+        "            wavfile.write(out_fpath, rate=out_sr, data=(wav_seg * np.iinfo(np.int16).max).astype(np.int16))"
+      ],
+      "metadata": {
+        "id": "LtLBGhGCbYYh"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "whisper_size = \"large\"\n",
+        "whisper_model = whisper.load_model(whisper_size)"
+      ],
+      "metadata": {
+        "id": "--wS7X95b--m"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "### 请将下方的**linghua.wav**修改成自己的.wav文件名，路径./custom_character_voice/**linghua**/也可以改为自己的角色名\n"
+      ],
+      "metadata": {
+        "id": "0wAE5HRXcCQ_"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "split_long_audio(whisper_model, \"./linghua.wav\", \"./custom_character_voice/linghua/\")"
+      ],
+      "metadata": {
+        "id": "3f7ljJhCcEbd"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "!python short_audio_transcribe.py --languages \"C\" --whisper_size large"
+      ],
+      "metadata": {
+        "id": "rBJDPe3ccVrP"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "#### 处理完成后，可以打开\"./filelists/short_character_anno.list\"文件进行微调"
+      ],
+      "metadata": {
+        "id": "4pesbcMjcikn"
+      }
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "### 2. 文本处理"
+      ],
+      "metadata": {
+        "id": "9pxo4KL-ceGI"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "!python preprocess_text.py"
+      ],
+      "metadata": {
+        "id": "_xfO2r_0cgCT"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "### 3. 运行bert_gen.py"
+      ],
+      "metadata": {
+        "id": "DoDs7lL6cu01"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "!python bert_gen.py"
+      ],
+      "metadata": {
+        "id": "jyiT28B3cxWX"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "### 4. 训练"
+      ],
+      "metadata": {
+        "id": "dHQPDFdbc04g"
+      }
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "#### 可以在\"./configs/config.json\"更改训练参数，包括epoch,学习率等"
+      ],
+      "metadata": {
+        "id": "gHNws-IUc6Sd"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "cd monotonic_align"
+      ],
+      "metadata": {
+        "id": "S56s0emH8BqN"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "!python setup.py build_ext --inplace"
+      ],
+      "metadata": {
+        "id": "6rLsyel-8KKc"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "cd .."
+      ],
+      "metadata": {
+        "id": "mUgA6ho2-XAN"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "#### 若为首次训练，请运行："
+      ],
+      "metadata": {
+        "id": "8vJ6VF__dCYW"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "!python train_ms.py -c ./configs/config.json"
+      ],
+      "metadata": {
+        "id": "iwHCWVijc5h6"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "#### 若为继续训练，请运行："
+      ],
+      "metadata": {
+        "id": "skAGULw2dKXW"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "!python train_ms.py -c ./configs/config.json --cont"
+      ],
+      "metadata": {
+        "id": "Ru09Gmavc2t4"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "### 5. 推理"
+      ],
+      "metadata": {
+        "id": "IinmucfadVLU"
+      }
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "#### 请将下方的**G_lastest.pth**修改为最新的模型文件，如**G_3400.pth**"
+      ],
+      "metadata": {
+        "id": "psBRLH_TdZDb"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "!python inference_webui.py --model_dir ./logs/OUTPUT_MODEL/G_latest.pth"
+      ],
+      "metadata": {
+        "id": "lOWVtUgMdUZa"
+      },
+      "execution_count": null,
+      "outputs": []
+    }
+  ]
+}

LICENSE

@@ -0,0 +1,674 @@
+                    GNU GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU General Public License is a free, copyleft license for
+software and other kinds of works.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+the GNU General Public License is intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.  We, the Free Software Foundation, use the
+GNU General Public License for most of our software; it applies also to
+any other work released this way by its authors.  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  To protect your rights, we need to prevent others from denying you
+these rights or asking you to surrender the rights.  Therefore, you have
+certain responsibilities if you distribute copies of the software, or if
+you modify it: responsibilities to respect the freedom of others.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must pass on to the recipients the same
+freedoms that you received.  You must make sure that they, too, receive
+or can get the source code.  And you must show them these terms so they
+know their rights.
+
+  Developers that use the GNU GPL protect your rights with two steps:
+(1) assert copyright on the software, and (2) offer you this License
+giving you legal permission to copy, distribute and/or modify it.
+
+  For the developers' and authors' protection, the GPL clearly explains
+that there is no warranty for this free software.  For both users' and
+authors' sake, the GPL requires that modified versions be marked as
+changed, so that their problems will not be attributed erroneously to
+authors of previous versions.
+
+  Some devices are designed to deny users access to install or run
+modified versions of the software inside them, although the manufacturer
+can do so.  This is fundamentally incompatible with the aim of
+protecting users' freedom to change the software.  The systematic
+pattern of such abuse occurs in the area of products for individuals to
+use, which is precisely where it is most unacceptable.  Therefore, we
+have designed this version of the GPL to prohibit the practice for those
+products.  If such problems arise substantially in other domains, we
+stand ready to extend this provision to those domains in future versions
+of the GPL, as needed to protect the freedom of users.
+
+  Finally, every program is threatened constantly by software patents.
+States should not allow patents to restrict development and use of
+software on general-purpose computers, but in those that do, we wish to
+avoid the special danger that patents applied to a free program could
+make it effectively proprietary.  To prevent this, the GPL assures that
+patents cannot be used to render the program non-free.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Use with the GNU Affero General Public License.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU Affero General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the special requirements of the GNU Affero General Public License,
+section 13, concerning interaction through a network will apply to the
+combination as such.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If the program does terminal interaction, make it output a short
+notice like this when it starts in an interactive mode:
+
+    <program>  Copyright (C) <year>  <name of author>
+    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, your program's commands
+might be different; for a GUI interface, you would use an "about box".
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU GPL, see
+<https://www.gnu.org/licenses/>.
+
+  The GNU General Public License does not permit incorporating your program
+into proprietary programs.  If your program is a subroutine library, you
+may consider it more useful to permit linking proprietary applications with
+the library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.  But first, please read
+<https://www.gnu.org/licenses/why-not-lgpl.html>.

README.md

@@ -1,13 +1,17 @@
----
-title: Bert VITS2
-emoji: 🏆
-colorFrom: blue
-colorTo: yellow
-sdk: gradio
-sdk_version: 3.44.3
-app_file: app.py
-pinned: false
-license: mit
----
+# Bert-VITS2-auto-preprocessing ⚡
+### [简体中文](https://github.com/KevinWang676/Bert-VITS2-quick-start/blob/main/README_zh.md)
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+This project is aimed at combining the features of [VITS-fast-fine-tuning](https://github.com/Plachtaa/VITS-fast-fine-tuning) with the original [Bert-VITS2](https://github.com/fishaudio/Bert-VITS2). Users can slice and label the uploaded audios automatically using the code we provide.
+
+```
+git clone https://github.com/KevinWang676/Bert-VITS2-quick-start.git
+cd Bert-VITS2-quick-start
+pip install -r requirements.txt
+```
+
+## References
++ [anyvoiceai/MassTTS](https://github.com/anyvoiceai/MassTTS)
++ [jaywalnut310/vits](https://github.com/jaywalnut310/vits)
++ [p0p4k/vits2_pytorch](https://github.com/p0p4k/vits2_pytorch)
++ [svc-develop-team/so-vits-svc](https://github.com/svc-develop-team/so-vits-svc)
++ [PaddlePaddle/PaddleSpeech](https://github.com/PaddlePaddle/PaddleSpeech)

README_zh.md

@@ -0,0 +1 @@
+

attentions.py

@@ -0,0 +1,343 @@
+import copy
+import math
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+import commons
+import modules
+from torch.nn.utils import weight_norm, remove_weight_norm
+class LayerNorm(nn.Module):
+  def __init__(self, channels, eps=1e-5):
+    super().__init__()
+    self.channels = channels
+    self.eps = eps
+
+    self.gamma = nn.Parameter(torch.ones(channels))
+    self.beta = nn.Parameter(torch.zeros(channels))
+
+  def forward(self, x):
+    x = x.transpose(1, -1)
+    x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+    return x.transpose(1, -1)
+
+ 
+   
+@torch.jit.script
+def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
+  n_channels_int = n_channels[0]
+  in_act = input_a + input_b
+  t_act = torch.tanh(in_act[:, :n_channels_int, :])
+  s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
+  acts = t_act * s_act
+  return acts
+
+class Encoder(nn.Module):
+   def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, isflow = True, **kwargs):
+    super().__init__()
+    self.hidden_channels = hidden_channels
+    self.filter_channels = filter_channels
+    self.n_heads = n_heads
+    self.n_layers = n_layers
+    self.kernel_size = kernel_size
+    self.p_dropout = p_dropout
+    self.window_size = window_size
+    if isflow:
+      cond_layer = torch.nn.Conv1d(256, 2*hidden_channels*n_layers, 1)
+      self.cond_pre = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, 1)
+      self.cond_layer = weight_norm(cond_layer, name='weight')
+      self.gin_channels = 256
+    self.cond_layer_idx = self.n_layers
+    if 'gin_channels' in kwargs:
+      self.gin_channels = kwargs['gin_channels']
+      if self.gin_channels != 0:
+        self.spk_emb_linear = nn.Linear(self.gin_channels, self.hidden_channels)
+        # vits2 says 3rd block, so idx is 2 by default
+        self.cond_layer_idx = kwargs['cond_layer_idx'] if 'cond_layer_idx' in kwargs else 2
+        print(self.gin_channels, self.cond_layer_idx)
+        assert self.cond_layer_idx < self.n_layers, 'cond_layer_idx should be less than n_layers'
+    self.drop = nn.Dropout(p_dropout)
+    self.attn_layers = nn.ModuleList()
+    self.norm_layers_1 = nn.ModuleList()
+    self.ffn_layers = nn.ModuleList()
+    self.norm_layers_2 = nn.ModuleList()
+    for i in range(self.n_layers):
+      self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
+      self.norm_layers_1.append(LayerNorm(hidden_channels))
+      self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))
+      self.norm_layers_2.append(LayerNorm(hidden_channels))
+   def forward(self, x, x_mask, g=None):
+    attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+    x = x * x_mask
+    for i in range(self.n_layers):
+      if i == self.cond_layer_idx and g is not None:
+        g = self.spk_emb_linear(g.transpose(1, 2))
+        g = g.transpose(1, 2)
+        x = x + g
+        x = x * x_mask
+      y = self.attn_layers[i](x, x, attn_mask)
+      y = self.drop(y)
+      x = self.norm_layers_1[i](x + y)
+
+      y = self.ffn_layers[i](x, x_mask)
+      y = self.drop(y)
+      x = self.norm_layers_2[i](x + y)
+    x = x * x_mask
+    return x
+
+
+class Decoder(nn.Module):
+  def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):
+    super().__init__()
+    self.hidden_channels = hidden_channels
+    self.filter_channels = filter_channels
+    self.n_heads = n_heads
+    self.n_layers = n_layers
+    self.kernel_size = kernel_size
+    self.p_dropout = p_dropout
+    self.proximal_bias = proximal_bias
+    self.proximal_init = proximal_init
+
+    self.drop = nn.Dropout(p_dropout)
+    self.self_attn_layers = nn.ModuleList()
+    self.norm_layers_0 = nn.ModuleList()
+    self.encdec_attn_layers = nn.ModuleList()
+    self.norm_layers_1 = nn.ModuleList()
+    self.ffn_layers = nn.ModuleList()
+    self.norm_layers_2 = nn.ModuleList()
+    for i in range(self.n_layers):
+      self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))
+      self.norm_layers_0.append(LayerNorm(hidden_channels))
+      self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
+      self.norm_layers_1.append(LayerNorm(hidden_channels))
+      self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
+      self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+  def forward(self, x, x_mask, h, h_mask):
+    """
+    x: decoder input
+    h: encoder output
+    """
+    self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)
+    encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+    x = x * x_mask
+    for i in range(self.n_layers):
+      y = self.self_attn_layers[i](x, x, self_attn_mask)
+      y = self.drop(y)
+      x = self.norm_layers_0[i](x + y)
+
+      y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
+      y = self.drop(y)
+      x = self.norm_layers_1[i](x + y)
+      
+      y = self.ffn_layers[i](x, x_mask)
+      y = self.drop(y)
+      x = self.norm_layers_2[i](x + y)
+    x = x * x_mask
+    return x
+
+
+class MultiHeadAttention(nn.Module):
+  def __init__(self, channels, out_channels, n_heads, p_dropout=0., window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False):
+    super().__init__()
+    assert channels % n_heads == 0
+
+    self.channels = channels
+    self.out_channels = out_channels
+    self.n_heads = n_heads
+    self.p_dropout = p_dropout
+    self.window_size = window_size
+    self.heads_share = heads_share
+    self.block_length = block_length
+    self.proximal_bias = proximal_bias
+    self.proximal_init = proximal_init
+    self.attn = None
+
+    self.k_channels = channels // n_heads
+    self.conv_q = nn.Conv1d(channels, channels, 1)
+    self.conv_k = nn.Conv1d(channels, channels, 1)
+    self.conv_v = nn.Conv1d(channels, channels, 1)
+    self.conv_o = nn.Conv1d(channels, out_channels, 1)
+    self.drop = nn.Dropout(p_dropout)
+
+    if window_size is not None:
+      n_heads_rel = 1 if heads_share else n_heads
+      rel_stddev = self.k_channels**-0.5
+      self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
+      self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
+
+    nn.init.xavier_uniform_(self.conv_q.weight)
+    nn.init.xavier_uniform_(self.conv_k.weight)
+    nn.init.xavier_uniform_(self.conv_v.weight)
+    if proximal_init:
+      with torch.no_grad():
+        self.conv_k.weight.copy_(self.conv_q.weight)
+        self.conv_k.bias.copy_(self.conv_q.bias)
+      
+  def forward(self, x, c, attn_mask=None):
+    q = self.conv_q(x)
+    k = self.conv_k(c)
+    v = self.conv_v(c)
+    
+    x, self.attn = self.attention(q, k, v, mask=attn_mask)
+
+    x = self.conv_o(x)
+    return x
+
+  def attention(self, query, key, value, mask=None):
+    # reshape [b, d, t] -> [b, n_h, t, d_k]
+    b, d, t_s, t_t = (*key.size(), query.size(2))
+    query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+    key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+    value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+
+    scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
+    if self.window_size is not None:
+      assert t_s == t_t, "Relative attention is only available for self-attention."
+      key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+      rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings)
+      scores_local = self._relative_position_to_absolute_position(rel_logits)
+      scores = scores + scores_local
+    if self.proximal_bias:
+      assert t_s == t_t, "Proximal bias is only available for self-attention."
+      scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
+    if mask is not None:
+      scores = scores.masked_fill(mask == 0, -1e4)
+      if self.block_length is not None:
+        assert t_s == t_t, "Local attention is only available for self-attention."
+        block_mask = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)
+        scores = scores.masked_fill(block_mask == 0, -1e4)
+    p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]
+    p_attn = self.drop(p_attn)
+    output = torch.matmul(p_attn, value)
+    if self.window_size is not None:
+      relative_weights = self._absolute_position_to_relative_position(p_attn)
+      value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
+      output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
+    output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t]
+    return output, p_attn
+
+  def _matmul_with_relative_values(self, x, y):
+    """
+    x: [b, h, l, m]
+    y: [h or 1, m, d]
+    ret: [b, h, l, d]
+    """
+    ret = torch.matmul(x, y.unsqueeze(0))
+    return ret
+
+  def _matmul_with_relative_keys(self, x, y):
+    """
+    x: [b, h, l, d]
+    y: [h or 1, m, d]
+    ret: [b, h, l, m]
+    """
+    ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+    return ret
+
+  def _get_relative_embeddings(self, relative_embeddings, length):
+    max_relative_position = 2 * self.window_size + 1
+    # Pad first before slice to avoid using cond ops.
+    pad_length = max(length - (self.window_size + 1), 0)
+    slice_start_position = max((self.window_size + 1) - length, 0)
+    slice_end_position = slice_start_position + 2 * length - 1
+    if pad_length > 0:
+      padded_relative_embeddings = F.pad(
+          relative_embeddings,
+          commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
+    else:
+      padded_relative_embeddings = relative_embeddings
+    used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]
+    return used_relative_embeddings
+
+  def _relative_position_to_absolute_position(self, x):
+    """
+    x: [b, h, l, 2*l-1]
+    ret: [b, h, l, l]
+    """
+    batch, heads, length, _ = x.size()
+    # Concat columns of pad to shift from relative to absolute indexing.
+    x = F.pad(x, commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))
+
+    # Concat extra elements so to add up to shape (len+1, 2*len-1).
+    x_flat = x.view([batch, heads, length * 2 * length])
+    x_flat = F.pad(x_flat, commons.convert_pad_shape([[0,0],[0,0],[0,length-1]]))
+
+    # Reshape and slice out the padded elements.
+    x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
+    return x_final
+
+  def _absolute_position_to_relative_position(self, x):
+    """
+    x: [b, h, l, l]
+    ret: [b, h, l, 2*l-1]
+    """
+    batch, heads, length, _ = x.size()
+    # padd along column
+    x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))
+    x_flat = x.view([batch, heads, length**2 + length*(length -1)])
+    # add 0's in the beginning that will skew the elements after reshape
+    x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
+    x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]
+    return x_final
+
+  def _attention_bias_proximal(self, length):
+    """Bias for self-attention to encourage attention to close positions.
+    Args:
+      length: an integer scalar.
+    Returns:
+      a Tensor with shape [1, 1, length, length]
+    """
+    r = torch.arange(length, dtype=torch.float32)
+    diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
+    return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
+
+
+class FFN(nn.Module):
+  def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):
+    super().__init__()
+    self.in_channels = in_channels
+    self.out_channels = out_channels
+    self.filter_channels = filter_channels
+    self.kernel_size = kernel_size
+    self.p_dropout = p_dropout
+    self.activation = activation
+    self.causal = causal
+
+    if causal:
+      self.padding = self._causal_padding
+    else:
+      self.padding = self._same_padding
+
+    self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
+    self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
+    self.drop = nn.Dropout(p_dropout)
+
+  def forward(self, x, x_mask):
+    x = self.conv_1(self.padding(x * x_mask))
+    if self.activation == "gelu":
+      x = x * torch.sigmoid(1.702 * x)
+    else:
+      x = torch.relu(x)
+    x = self.drop(x)
+    x = self.conv_2(self.padding(x * x_mask))
+    return x * x_mask
+  
+  def _causal_padding(self, x):
+    if self.kernel_size == 1:
+      return x
+    pad_l = self.kernel_size - 1
+    pad_r = 0
+    padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+    x = F.pad(x, commons.convert_pad_shape(padding))
+    return x
+
+  def _same_padding(self, x):
+    if self.kernel_size == 1:
+      return x
+    pad_l = (self.kernel_size - 1) // 2
+    pad_r = self.kernel_size // 2
+    padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+    x = F.pad(x, commons.convert_pad_shape(padding))
+    return x

bert/chinese-roberta-wwm-ext-large/.gitattributes

@@ -0,0 +1,9 @@
+*.bin.* filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tar.gz filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text

bert/chinese-roberta-wwm-ext-large/README.md

@@ -0,0 +1,64 @@
+---
+language: 
+- zh
+tags:
+- bert
+license: "apache-2.0"
+---
+
+How to download the files:
+```
+curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh | sudo bash
+sudo apt-get install git-lfs
+git clone https://huggingface.co/hfl/chinese-roberta-wwm-ext-large
+```
+
+# Please use 'Bert' related functions to load this model!
+
+## Chinese BERT with Whole Word Masking
+For further accelerating Chinese natural language processing, we provide **Chinese pre-trained BERT with Whole Word Masking**. 
+
+**[Pre-Training with Whole Word Masking for Chinese BERT](https://arxiv.org/abs/1906.08101)**  
+Yiming Cui, Wanxiang Che, Ting Liu, Bing Qin, Ziqing Yang, Shijin Wang, Guoping Hu
+
+This repository is developed based on：https://github.com/google-research/bert
+
+You may also interested in,
+- Chinese BERT series: https://github.com/ymcui/Chinese-BERT-wwm
+- Chinese MacBERT: https://github.com/ymcui/MacBERT
+- Chinese ELECTRA: https://github.com/ymcui/Chinese-ELECTRA
+- Chinese XLNet: https://github.com/ymcui/Chinese-XLNet
+- Knowledge Distillation Toolkit - TextBrewer: https://github.com/airaria/TextBrewer
+
+More resources by HFL: https://github.com/ymcui/HFL-Anthology
+
+## Citation
+If you find the technical report or resource is useful, please cite the following technical report in your paper.
+- Primary: https://arxiv.org/abs/2004.13922
+```
+@inproceedings{cui-etal-2020-revisiting,
+    title = "Revisiting Pre-Trained Models for {C}hinese Natural Language Processing",
+    author = "Cui, Yiming  and
+      Che, Wanxiang  and
+      Liu, Ting  and
+      Qin, Bing  and
+      Wang, Shijin  and
+      Hu, Guoping",
+    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings",
+    month = nov,
+    year = "2020",
+    address = "Online",
+    publisher = "Association for Computational Linguistics",
+    url = "https://www.aclweb.org/anthology/2020.findings-emnlp.58",
+    pages = "657--668",
+}
+```
+- Secondary: https://arxiv.org/abs/1906.08101  
+```
+@article{chinese-bert-wwm,
+  title={Pre-Training with Whole Word Masking for Chinese BERT},
+  author={Cui, Yiming and Che, Wanxiang and Liu, Ting and Qin, Bing and Yang, Ziqing and Wang, Shijin and Hu, Guoping},
+  journal={arXiv preprint arXiv:1906.08101},
+  year={2019}
+ }
+```

bert/chinese-roberta-wwm-ext-large/added_tokens.json

@@ -0,0 +1 @@
+{}

bert/chinese-roberta-wwm-ext-large/config.json

@@ -0,0 +1,28 @@
+{
+  "architectures": [
+    "BertForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "directionality": "bidi",
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 24,
+  "output_past": true,
+  "pad_token_id": 0,
+  "pooler_fc_size": 768,
+  "pooler_num_attention_heads": 12,
+  "pooler_num_fc_layers": 3,
+  "pooler_size_per_head": 128,
+  "pooler_type": "first_token_transform",
+  "type_vocab_size": 2,
+  "vocab_size": 21128
+}

bert/chinese-roberta-wwm-ext-large/special_tokens_map.json

@@ -0,0 +1 @@
+{"unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]"}

bert/chinese-roberta-wwm-ext-large/tokenizer_config.json

@@ -0,0 +1 @@
+{"init_inputs": []}

bert_gen.py

@@ -0,0 +1,53 @@
+import torch
+from torch.utils.data import DataLoader
+from multiprocessing import Pool
+import commons
+import utils
+from data_utils import TextAudioSpeakerLoader, TextAudioSpeakerCollate
+from tqdm import tqdm
+import warnings
+
+from text import cleaned_text_to_sequence, get_bert
+
+config_path = 'configs/config.json'
+hps = utils.get_hparams_from_file(config_path)
+
+def process_line(line):
+    _id, spk, language_str, text, phones, tone, word2ph = line.strip().split("|")
+    phone = phones.split(" ")
+    tone = [int(i) for i in tone.split(" ")]
+    word2ph = [int(i) for i in word2ph.split(" ")]
+    w2pho = [i for i in word2ph]
+    word2ph = [i for i in word2ph]
+    phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str)
+
+    if hps.data.add_blank:
+        phone = commons.intersperse(phone, 0)
+        tone = commons.intersperse(tone, 0)
+        language = commons.intersperse(language, 0)
+        for i in range(len(word2ph)):
+            word2ph[i] = word2ph[i] * 2
+        word2ph[0] += 1
+    wav_path = f'{_id}'
+
+    bert_path = wav_path.replace(".wav", ".bert.pt")
+    try:
+        bert = torch.load(bert_path)
+        assert bert.shape[-1] == len(phone)
+    except:
+        bert = get_bert(text, word2ph, language_str)
+        assert bert.shape[-1] == len(phone)
+        torch.save(bert, bert_path)
+
+
+if __name__ == '__main__':
+    lines = []
+    with open(hps.data.training_files, encoding='utf-8' ) as f:
+        lines.extend(f.readlines())
+
+   # with open(hps.data.validation_files, encoding='utf-8' ) as f:
+       # lines.extend(f.readlines())
+
+    with Pool(processes=2) as pool: #A100 40GB suitable config,if coom,please decrease the processess number.
+        for _ in tqdm(pool.imap_unordered(process_line, lines)):
+            pass

commons.py

@@ -0,0 +1,161 @@
+import math
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+
+def init_weights(m, mean=0.0, std=0.01):
+  classname = m.__class__.__name__
+  if classname.find("Conv") != -1:
+    m.weight.data.normal_(mean, std)
+
+
+def get_padding(kernel_size, dilation=1):
+  return int((kernel_size*dilation - dilation)/2)
+
+
+def convert_pad_shape(pad_shape):
+  l = pad_shape[::-1]
+  pad_shape = [item for sublist in l for item in sublist]
+  return pad_shape
+
+
+def intersperse(lst, item):
+  result = [item] * (len(lst) * 2 + 1)
+  result[1::2] = lst
+  return result
+
+
+def kl_divergence(m_p, logs_p, m_q, logs_q):
+  """KL(P||Q)"""
+  kl = (logs_q - logs_p) - 0.5
+  kl += 0.5 * (torch.exp(2. * logs_p) + ((m_p - m_q)**2)) * torch.exp(-2. * logs_q)
+  return kl
+
+
+def rand_gumbel(shape):
+  """Sample from the Gumbel distribution, protect from overflows."""
+  uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
+  return -torch.log(-torch.log(uniform_samples))
+
+
+def rand_gumbel_like(x):
+  g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
+  return g
+
+
+def slice_segments(x, ids_str, segment_size=4):
+  ret = torch.zeros_like(x[:, :, :segment_size])
+  for i in range(x.size(0)):
+    idx_str = ids_str[i]
+    idx_end = idx_str + segment_size
+    ret[i] = x[i, :, idx_str:idx_end]
+  return ret
+
+
+def rand_slice_segments(x, x_lengths=None, segment_size=4):
+  b, d, t = x.size()
+  if x_lengths is None:
+    x_lengths = t
+  ids_str_max = x_lengths - segment_size + 1
+  ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
+  ret = slice_segments(x, ids_str, segment_size)
+  return ret, ids_str
+
+
+def get_timing_signal_1d(
+    length, channels, min_timescale=1.0, max_timescale=1.0e4):
+  position = torch.arange(length, dtype=torch.float)
+  num_timescales = channels // 2
+  log_timescale_increment = (
+      math.log(float(max_timescale) / float(min_timescale)) /
+      (num_timescales - 1))
+  inv_timescales = min_timescale * torch.exp(
+      torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment)
+  scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
+  signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
+  signal = F.pad(signal, [0, 0, 0, channels % 2])
+  signal = signal.view(1, channels, length)
+  return signal
+
+
+def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
+  b, channels, length = x.size()
+  signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+  return x + signal.to(dtype=x.dtype, device=x.device)
+
+
+def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
+  b, channels, length = x.size()
+  signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+  return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
+
+
+def subsequent_mask(length):
+  mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
+  return mask
+
+
+@torch.jit.script
+def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
+  n_channels_int = n_channels[0]
+  in_act = input_a + input_b
+  t_act = torch.tanh(in_act[:, :n_channels_int, :])
+  s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
+  acts = t_act * s_act
+  return acts
+
+
+def convert_pad_shape(pad_shape):
+  l = pad_shape[::-1]
+  pad_shape = [item for sublist in l for item in sublist]
+  return pad_shape
+
+
+def shift_1d(x):
+  x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
+  return x
+
+
+def sequence_mask(length, max_length=None):
+  if max_length is None:
+    max_length = length.max()
+  x = torch.arange(max_length, dtype=length.dtype, device=length.device)
+  return x.unsqueeze(0) < length.unsqueeze(1)
+
+
+def generate_path(duration, mask):
+  """
+  duration: [b, 1, t_x]
+  mask: [b, 1, t_y, t_x]
+  """
+  device = duration.device
+  
+  b, _, t_y, t_x = mask.shape
+  cum_duration = torch.cumsum(duration, -1)
+  
+  cum_duration_flat = cum_duration.view(b * t_x)
+  path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
+  path = path.view(b, t_x, t_y)
+  path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
+  path = path.unsqueeze(1).transpose(2,3) * mask
+  return path
+
+
+def clip_grad_value_(parameters, clip_value, norm_type=2):
+  if isinstance(parameters, torch.Tensor):
+    parameters = [parameters]
+  parameters = list(filter(lambda p: p.grad is not None, parameters))
+  norm_type = float(norm_type)
+  if clip_value is not None:
+    clip_value = float(clip_value)
+
+  total_norm = 0
+  for p in parameters:
+    param_norm = p.grad.data.norm(norm_type)
+    total_norm += param_norm.item() ** norm_type
+    if clip_value is not None:
+      p.grad.data.clamp_(min=-clip_value, max=clip_value)
+  total_norm = total_norm ** (1. / norm_type)
+  return total_norm

configs/config.json

@@ -0,0 +1,342 @@
+{
+  "train": {
+    "log_interval": 10,
+    "eval_interval": 100,
+    "seed": 52,
+    "epochs": 1000,
+    "learning_rate": 0.00015,
+    "betas": [
+      0.8,
+      0.99
+    ],
+    "eps": 1e-09,
+    "batch_size": 12,
+    "fp16_run": false,
+    "lr_decay": 0.999875,
+    "segment_size": 16384,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0
+  },
+  "data": {
+    "use_mel_posterior_encoder": false,
+    "training_files": "filelists/train.list",
+    "validation_files": "filelists/val.list",
+    "max_wav_value": 32768.0,
+    "sampling_rate": 44100,
+    "filter_length": 2048,
+    "hop_length": 512,
+    "win_length": 2048,
+    "n_mel_channels": 128,
+    "mel_fmin": 0.0,
+    "mel_fmax": null,
+    "add_blank": true,
+    "n_speakers": 256,
+    "cleaned_text": true,
+    "spk2id": {
+      "丹恒": 0,
+      "克拉拉": 1,
+      "穹": 2,
+      "「信使」": 3,
+      "史瓦罗": 4,
+      "彦卿": 5,
+      "晴霓": 6,
+      "杰帕德": 7,
+      "素裳": 8,
+      "绿芙蓉": 9,
+      "罗刹": 10,
+      "艾丝妲": 11,
+      "黑塔": 12,
+      "丹枢": 13,
+      "希露瓦": 14,
+      "白露": 15,
+      "费斯曼": 16,
+      "停云": 17,
+      "可可利亚": 18,
+      "景元": 19,
+      "螺丝咕姆": 20,
+      "青镞": 21,
+      "公输师傅": 22,
+      "卡芙卡": 23,
+      "大毫": 24,
+      "驭空": 25,
+      "半夏": 26,
+      "奥列格": 27,
+      "娜塔莎": 28,
+      "桑博": 29,
+      "瓦尔特": 30,
+      "阿兰": 31,
+      "伦纳德": 32,
+      "佩拉": 33,
+      "卡波特": 34,
+      "帕姆": 35,
+      "帕斯卡": 36,
+      "青雀": 37,
+      "三月七": 38,
+      "刃": 39,
+      "姬子": 40,
+      "布洛妮娅": 41,
+      "希儿": 42,
+      "星": 43,
+      "符玄": 44,
+      "虎克": 45,
+      "银狼": 46,
+      "镜流": 47,
+      "「博士」": 48,
+      "「大肉丸」": 49,
+      "九条裟罗": 50,
+      "佐西摩斯": 51,
+      "刻晴": 52,
+      "博易": 53,
+      "卡维": 54,
+      "可莉": 55,
+      "嘉玛": 56,
+      "埃舍尔": 57,
+      "塔杰·拉德卡尼": 58,
+      "大慈树王": 59,
+      "宵宫": 60,
+      "康纳": 61,
+      "影": 62,
+      "枫原万叶": 63,
+      "欧菲妮": 64,
+      "玛乔丽": 65,
+      "珊瑚": 66,
+      "田铁嘴": 67,
+      "砂糖": 68,
+      "神里绫华": 69,
+      "罗莎莉亚": 70,
+      "荒泷一斗": 71,
+      "莎拉": 72,
+      "迪希雅": 73,
+      "钟离": 74,
+      "阿圆": 75,
+      "阿娜耶": 76,
+      "阿拉夫": 77,
+      "雷泽": 78,
+      "香菱": 79,
+      "龙二": 80,
+      "「公子」": 81,
+      "「白老先生」": 82,
+      "优菈": 83,
+      "凯瑟琳": 84,
+      "哲平": 85,
+      "夏洛蒂": 86,
+      "安柏": 87,
+      "巴达维": 88,
+      "式大将": 89,
+      "斯坦利": 90,
+      "毗伽尔": 91,
+      "海妮耶": 92,
+      "爱德琳": 93,
+      "纳西妲": 94,
+      "老孟": 95,
+      "芙宁娜": 96,
+      "阿守": 97,
+      "阿祇": 98,
+      "丹吉尔": 99,
+      "丽莎": 100,
+      "五郎": 101,
+      "元太": 102,
+      "克列门特": 103,
+      "克罗索": 104,
+      "北斗": 105,
+      "埃勒曼": 106,
+      "天目十五": 107,
+      "奥兹": 108,
+      "恶龙": 109,
+      "早柚": 110,
+      "杜拉夫": 111,
+      "松浦": 112,
+      "柊千里": 113,
+      "甘雨": 114,
+      "石头": 115,
+      "纯水精灵？": 116,
+      "羽生田千鹤": 117,
+      "莱依拉": 118,
+      "菲谢尔": 119,
+      "言笑": 120,
+      "诺艾尔": 121,
+      "赛诺": 122,
+      "辛焱": 123,
+      "迪娜泽黛": 124,
+      "那维莱特": 125,
+      "八重神子": 126,
+      "凯亚": 127,
+      "吴船长": 128,
+      "埃德": 129,
+      "天叔": 130,
+      "女士": 131,
+      "恕筠": 132,
+      "提纳里": 133,
+      "派蒙": 134,
+      "流浪者": 135,
+      "深渊使徒": 136,
+      "玛格丽特": 137,
+      "珐露珊": 138,
+      "琴": 139,
+      "瑶瑶": 140,
+      "留云借风真君": 141,
+      "绮良良": 142,
+      "舒伯特": 143,
+      "荧": 144,
+      "莫娜": 145,
+      "行秋": 146,
+      "迈勒斯": 147,
+      "阿佩普": 148,
+      "鹿野奈奈": 149,
+      "七七": 150,
+      "伊迪娅": 151,
+      "博来": 152,
+      "坎蒂丝": 153,
+      "埃尔欣根": 154,
+      "埃泽": 155,
+      "塞琉斯": 156,
+      "夜兰": 157,
+      "常九爷": 158,
+      "悦": 159,
+      "戴因斯雷布": 160,
+      "笼钓瓶一心": 161,
+      "纳比尔": 162,
+      "胡桃": 163,
+      "艾尔海森": 164,
+      "艾莉丝": 165,
+      "菲米尼": 166,
+      "蒂玛乌斯": 167,
+      "迪奥娜": 168,
+      "阿晃": 169,
+      "阿洛瓦": 170,
+      "陆行岩本真蕈·元素生命": 171,
+      "雷电将军": 172,
+      "魈": 173,
+      "鹿野院平藏": 174,
+      "「女士」": 175,
+      "「散兵」": 176,
+      "凝光": 177,
+      "妮露": 178,
+      "娜维娅": 179,
+      "宛烟": 180,
+      "慧心": 181,
+      "托克": 182,
+      "托马": 183,
+      "掇星攫辰天君": 184,
+      "旁白": 185,
+      "浮游水蕈兽·元素生命": 186,
+      "烟绯": 187,
+      "玛塞勒": 188,
+      "百闻": 189,
+      "知易": 190,
+      "米卡": 191,
+      "西拉杰": 192,
+      "迪卢克": 193,
+      "重云": 194,
+      "阿扎尔": 195,
+      "霍夫曼": 196,
+      "上杉": 197,
+      "久利须": 198,
+      "嘉良": 199,
+      "回声海螺": 200,
+      "多莉": 201,
+      "安西": 202,
+      "德沃沙克": 203,
+      "拉赫曼": 204,
+      "林尼": 205,
+      "查尔斯": 206,
+      "深渊法师": 207,
+      "温迪": 208,
+      "爱贝尔": 209,
+      "珊瑚宫心海": 210,
+      "班尼特": 211,
+      "琳妮特": 212,
+      "申鹤": 213,
+      "神里绫人": 214,
+      "艾伯特": 215,
+      "萍姥姥": 216,
+      "萨赫哈蒂": 217,
+      "萨齐因": 218,
+      "阿尔卡米": 219,
+      "阿贝多": 220,
+      "anzai": 221,
+      "久岐忍": 222,
+      "九条镰治": 223,
+      "云堇": 224,
+      "伊利亚斯": 225,
+      "埃洛伊": 226,
+      "塞塔蕾": 227,
+      "拉齐": 228,
+      "昆钧": 229,
+      "柯莱": 230,
+      "沙扎曼": 231,
+      "海芭夏": 232,
+      "白术": 233,
+      "空": 234,
+      "艾文": 235,
+      "芭芭拉": 236,
+      "莫塞伊思": 237,
+      "莺儿": 238,
+      "达达利亚": 239,
+      "迈蒙": 240,
+      "长生": 241,
+      "阿巴图伊": 242,
+      "陆景和": 243,
+      "莫弈": 244,
+      "夏彦": 245,
+      "左然": 246,
+      "标贝": 247
+    }
+  },
+  "model": {
+    "use_spk_conditioned_encoder": true,
+    "use_noise_scaled_mas": true,
+    "use_mel_posterior_encoder": false,
+    "use_duration_discriminator": true,
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "resblock": "1",
+    "resblock_kernel_sizes": [
+      3,
+      7,
+      11
+    ],
+    "resblock_dilation_sizes": [
+      [
+        1,
+        3,
+        5
+      ],
+      [
+        1,
+        3,
+        5
+      ],
+      [
+        1,
+        3,
+        5
+      ]
+    ],
+    "upsample_rates": [
+      8,
+      8,
+      2,
+      2,
+      2
+    ],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [
+      16,
+      16,
+      8,
+      2,
+      2
+    ],
+    "n_layers_q": 3,
+    "use_spectral_norm": false,
+    "gin_channels": 256
+  }
+}

custom_character_voice/44100.txt

@@ -0,0 +1 @@
+

data_utils.py

@@ -0,0 +1,332 @@
+import time
+import os
+import random
+import numpy as np
+import torch
+import torch.utils.data
+import torchaudio
+import commons
+from mel_processing import spectrogram_torch, mel_spectrogram_torch, spec_to_mel_torch
+from utils import load_wav_to_torch, load_filepaths_and_text
+from text import cleaned_text_to_sequence, get_bert
+
+"""Multi speaker version"""
+
+
+class TextAudioSpeakerLoader(torch.utils.data.Dataset):
+    """
+        1) loads audio, speaker_id, text pairs
+        2) normalizes text and converts them to sequences of integers
+        3) computes spectrograms from audio files.
+    """
+
+    def __init__(self, audiopaths_sid_text, hparams):
+        self.audiopaths_sid_text = load_filepaths_and_text(audiopaths_sid_text)
+        self.max_wav_value = hparams.max_wav_value
+        self.sampling_rate = hparams.sampling_rate
+        self.filter_length = hparams.filter_length
+        self.hop_length = hparams.hop_length
+        self.win_length = hparams.win_length
+        self.sampling_rate = hparams.sampling_rate
+        self.spk_map = hparams.spk2id
+        self.hparams = hparams
+
+        self.use_mel_spec_posterior = getattr(hparams, "use_mel_posterior_encoder", False)
+        if self.use_mel_spec_posterior:
+            self.n_mel_channels = getattr(hparams, "n_mel_channels", 80)
+
+        self.cleaned_text = getattr(hparams, "cleaned_text", False)
+
+        self.add_blank = hparams.add_blank
+        self.min_text_len = getattr(hparams, "min_text_len", 1)
+        self.max_text_len = getattr(hparams, "max_text_len", 300)
+
+        random.seed(1234)
+        random.shuffle(self.audiopaths_sid_text)
+        self._filter()
+
+    def _filter(self):
+        """
+        Filter text & store spec lengths
+        """
+        # Store spectrogram lengths for Bucketing
+        # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
+        # spec_length = wav_length // hop_length
+
+        audiopaths_sid_text_new = []
+        lengths = []
+        skipped = 0
+        for _id, spk, language, text, phones, tone, word2ph in self.audiopaths_sid_text:
+            audiopath = f'{_id}'
+            if self.min_text_len <= len(phones) and len(phones) <= self.max_text_len:
+                phones = phones.split(" ")
+                tone = [int(i) for i in tone.split(" ")]
+                word2ph = [int(i) for i in word2ph.split(" ")]
+                audiopaths_sid_text_new.append([audiopath, spk, language, text, phones, tone, word2ph])
+                lengths.append(os.path.getsize(audiopath) // (2 * self.hop_length))
+            else:
+                skipped += 1
+        print("skipped: ", skipped, ", total: ", len(self.audiopaths_sid_text))
+        self.audiopaths_sid_text = audiopaths_sid_text_new
+        self.lengths = lengths
+
+    def get_audio_text_speaker_pair(self, audiopath_sid_text):
+        # separate filename, speaker_id and text
+        audiopath, sid, language, text, phones, tone, word2ph = audiopath_sid_text
+
+        bert, phones, tone, language = self.get_text(text, word2ph, phones, tone, language, audiopath)
+
+        spec, wav = self.get_audio(audiopath)
+        sid = torch.LongTensor([int(self.spk_map[sid])])
+        return (phones, spec, wav, sid, tone, language, bert)
+
+    def get_audio(self, filename):
+        audio_norm, sampling_rate = torchaudio.load(filename, frame_offset=0, num_frames=-1, normalize=True, channels_first=True)
+        '''
+        audio, sampling_rate = load_wav_to_torch(filename)
+        if sampling_rate != self.sampling_rate:
+            raise ValueError("{} {} SR doesn't match target {} SR".format(
+                sampling_rate, self.sampling_rate))
+        audio_norm = audio / self.max_wav_value
+        audio_norm = audio_norm.unsqueeze(0)
+        '''
+        spec_filename = filename.replace(".wav", ".spec.pt")
+        if self.use_mel_spec_posterior:
+            spec_filename = spec_filename.replace(".spec.pt", ".mel.pt")
+        if os.path.exists(spec_filename):
+            spec = torch.load(spec_filename)
+        else:
+            if self.use_mel_spec_posterior:
+                # if os.path.exists(filename.replace(".wav", ".spec.pt")):
+                #     # spec, n_fft, num_mels, sampling_rate, fmin, fmax
+                #     spec = spec_to_mel_torch(
+                #         torch.load(filename.replace(".wav", ".spec.pt")), 
+                #         self.filter_length, self.n_mel_channels, self.sampling_rate,
+                #         self.hparams.mel_fmin, self.hparams.mel_fmax)
+                spec = mel_spectrogram_torch(audio_norm, self.filter_length,
+                    self.n_mel_channels, self.sampling_rate, self.hop_length,
+                    self.win_length, self.hparams.mel_fmin, self.hparams.mel_fmax, center=False)
+            else:
+                spec = spectrogram_torch(audio_norm, self.filter_length,
+                    self.sampling_rate, self.hop_length, self.win_length,
+                    center=False)
+            spec = torch.squeeze(spec, 0)
+            torch.save(spec, spec_filename)
+        return spec, audio_norm
+
+    def get_text(self, text, word2ph, phone, tone, language_str, wav_path):
+        # print(text, word2ph,phone, tone, language_str)
+        pold = phone
+        w2pho = [i for i in word2ph]
+        word2ph = [i for i in word2ph]
+        phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str)
+        pold2 = phone
+
+        if self.add_blank:
+            p1 = len(phone)
+            phone = commons.intersperse(phone, 0)
+            p2 = len(phone)
+            t1 = len(tone)
+            tone = commons.intersperse(tone, 0)
+            t2 = len(tone)
+            language = commons.intersperse(language, 0)
+            for i in range(len(word2ph)):
+                word2ph[i] = word2ph[i] * 2
+            word2ph[0] += 1
+        bert_path = wav_path.replace(".wav", ".bert.pt")
+        try:
+            bert = torch.load(bert_path)
+            assert bert.shape[-1] == len(phone)
+        except:
+            bert = get_bert(text, word2ph, language_str)
+            torch.save(bert, bert_path)
+            #print(bert.shape[-1], bert_path, text, pold)
+            assert bert.shape[-1] == len(phone)
+
+        assert bert.shape[-1] == len(phone), (
+        bert.shape, len(phone), sum(word2ph), p1, p2, t1, t2, pold, pold2, word2ph, text, w2pho)
+        phone = torch.LongTensor(phone)
+        tone = torch.LongTensor(tone)
+        language = torch.LongTensor(language)
+        return bert, phone, tone, language
+
+    def get_sid(self, sid):
+        sid = torch.LongTensor([int(sid)])
+        return sid
+
+    def __getitem__(self, index):
+        return self.get_audio_text_speaker_pair(self.audiopaths_sid_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_sid_text)
+
+
+class TextAudioSpeakerCollate():
+    """ Zero-pads model inputs and targets
+    """
+
+    def __init__(self, return_ids=False):
+        self.return_ids = return_ids
+
+    def __call__(self, batch):
+        """Collate's training batch from normalized text, audio and speaker identities
+        PARAMS
+        ------
+        batch: [text_normalized, spec_normalized, wav_normalized, sid]
+        """
+        # Right zero-pad all one-hot text sequences to max input length
+        _, ids_sorted_decreasing = torch.sort(
+            torch.LongTensor([x[1].size(1) for x in batch]),
+            dim=0, descending=True)
+
+        max_text_len = max([len(x[0]) for x in batch])
+        max_spec_len = max([x[1].size(1) for x in batch])
+        max_wav_len = max([x[2].size(1) for x in batch])
+
+        text_lengths = torch.LongTensor(len(batch))
+        spec_lengths = torch.LongTensor(len(batch))
+        wav_lengths = torch.LongTensor(len(batch))
+        sid = torch.LongTensor(len(batch))
+
+        text_padded = torch.LongTensor(len(batch), max_text_len)
+        tone_padded = torch.LongTensor(len(batch), max_text_len)
+        language_padded = torch.LongTensor(len(batch), max_text_len)
+        bert_padded = torch.FloatTensor(len(batch), 1024, max_text_len)
+
+        spec_padded = torch.FloatTensor(len(batch), batch[0][1].size(0), max_spec_len)
+        wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
+        text_padded.zero_()
+        tone_padded.zero_()
+        language_padded.zero_()
+        spec_padded.zero_()
+        wav_padded.zero_()
+        bert_padded.zero_()
+        for i in range(len(ids_sorted_decreasing)):
+            row = batch[ids_sorted_decreasing[i]]
+
+            text = row[0]
+            text_padded[i, :text.size(0)] = text
+            text_lengths[i] = text.size(0)
+
+            spec = row[1]
+            spec_padded[i, :, :spec.size(1)] = spec
+            spec_lengths[i] = spec.size(1)
+
+            wav = row[2]
+            wav_padded[i, :, :wav.size(1)] = wav
+            wav_lengths[i] = wav.size(1)
+
+            sid[i] = row[3]
+
+            tone = row[4]
+            tone_padded[i, :tone.size(0)] = tone
+
+            language = row[5]
+            language_padded[i, :language.size(0)] = language
+
+            bert = row[6]
+            bert_padded[i, :, :bert.size(1)] = bert
+
+        return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, sid, tone_padded, language_padded, bert_padded
+
+
+class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
+    """
+    Maintain similar input lengths in a batch.
+    Length groups are specified by boundaries.
+    Ex) boundaries = [b1, b2, b3] -> any batch is included either {x | b1 < length(x) <=b2} or {x | b2 < length(x) <= b3}.
+
+    It removes samples which are not included in the boundaries.
+    Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1 or length(x) > b3 are discarded.
+    """
+
+    def __init__(self, dataset, batch_size, boundaries, num_replicas=None, rank=None, shuffle=True):
+        super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
+        self.lengths = dataset.lengths
+        self.batch_size = batch_size
+        self.boundaries = boundaries
+
+        self.buckets, self.num_samples_per_bucket = self._create_buckets()
+        self.total_size = sum(self.num_samples_per_bucket)
+        self.num_samples = self.total_size // self.num_replicas
+
+    def _create_buckets(self):
+        buckets = [[] for _ in range(len(self.boundaries) - 1)]
+        for i in range(len(self.lengths)):
+            length = self.lengths[i]
+            idx_bucket = self._bisect(length)
+            if idx_bucket != -1:
+                buckets[idx_bucket].append(i)
+
+        for i in range(len(buckets) - 1, 0, -1):
+            if len(buckets[i]) == 0:
+                buckets.pop(i)
+                self.boundaries.pop(i + 1)
+
+        num_samples_per_bucket = []
+        for i in range(len(buckets)):
+            len_bucket = len(buckets[i])
+            total_batch_size = self.num_replicas * self.batch_size
+            rem = (total_batch_size - (len_bucket % total_batch_size)) % total_batch_size
+            num_samples_per_bucket.append(len_bucket + rem)
+        return buckets, num_samples_per_bucket
+
+    def __iter__(self):
+        # deterministically shuffle based on epoch
+        g = torch.Generator()
+        g.manual_seed(self.epoch)
+
+        indices = []
+        if self.shuffle:
+            for bucket in self.buckets:
+                indices.append(torch.randperm(len(bucket), generator=g).tolist())
+        else:
+            for bucket in self.buckets:
+                indices.append(list(range(len(bucket))))
+
+        batches = []
+        for i in range(len(self.buckets)):
+            bucket = self.buckets[i]
+            len_bucket = len(bucket)
+            if (len_bucket == 0):
+                continue
+            ids_bucket = indices[i]
+            num_samples_bucket = self.num_samples_per_bucket[i]
+
+            # add extra samples to make it evenly divisible
+            rem = num_samples_bucket - len_bucket
+            ids_bucket = ids_bucket + ids_bucket * (rem // len_bucket) + ids_bucket[:(rem % len_bucket)]
+
+            # subsample
+            ids_bucket = ids_bucket[self.rank::self.num_replicas]
+
+            # batching
+            for j in range(len(ids_bucket) // self.batch_size):
+                batch = [bucket[idx] for idx in ids_bucket[j * self.batch_size:(j + 1) * self.batch_size]]
+                batches.append(batch)
+
+        if self.shuffle:
+            batch_ids = torch.randperm(len(batches), generator=g).tolist()
+            batches = [batches[i] for i in batch_ids]
+        self.batches = batches
+
+        assert len(self.batches) * self.batch_size == self.num_samples
+        return iter(self.batches)
+
+    def _bisect(self, x, lo=0, hi=None):
+        if hi is None:
+            hi = len(self.boundaries) - 1
+
+        if hi > lo:
+            mid = (hi + lo) // 2
+            if self.boundaries[mid] < x and x <= self.boundaries[mid + 1]:
+                return mid
+            elif x <= self.boundaries[mid]:
+                return self._bisect(x, lo, mid)
+            else:
+                return self._bisect(x, mid + 1, hi)
+        else:
+            return -1
+
+    def __len__(self):
+        return self.num_samples // self.batch_size

filelists/esd.list

@@ -0,0 +1,2 @@
+Example:
+{wav_path}|{speaker_name}|{language}|{text}

inference_webui.py

@@ -0,0 +1,132 @@
+import sys, os
+
+if sys.platform == "darwin":
+    os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
+
+import logging
+
+logging.getLogger("numba").setLevel(logging.WARNING)
+logging.getLogger("markdown_it").setLevel(logging.WARNING)
+logging.getLogger("urllib3").setLevel(logging.WARNING)
+logging.getLogger("matplotlib").setLevel(logging.WARNING)
+
+logging.basicConfig(level=logging.INFO, format="| %(name)s | %(levelname)s | %(message)s")
+
+logger = logging.getLogger(__name__)
+
+import torch
+import argparse
+import commons
+import utils
+from models import SynthesizerTrn
+from text.symbols import symbols
+from text import cleaned_text_to_sequence, get_bert
+from text.cleaner import clean_text
+import gradio as gr
+import webbrowser
+
+
+net_g = None
+
+
+def get_text(text, language_str, hps):
+    norm_text, phone, tone, word2ph = clean_text(text, language_str)
+    phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str)
+
+    if hps.data.add_blank:
+        phone = commons.intersperse(phone, 0)
+        tone = commons.intersperse(tone, 0)
+        language = commons.intersperse(language, 0)
+        for i in range(len(word2ph)):
+            word2ph[i] = word2ph[i] * 2
+        word2ph[0] += 1
+    bert = get_bert(norm_text, word2ph, language_str)
+    del word2ph
+
+    assert bert.shape[-1] == len(phone)
+
+    phone = torch.LongTensor(phone)
+    tone = torch.LongTensor(tone)
+    language = torch.LongTensor(language)
+
+    return bert, phone, tone, language
+
+def infer(text, sdp_ratio, noise_scale, noise_scale_w, length_scale, sid):
+    global net_g
+    bert, phones, tones, lang_ids = get_text(text, "ZH", hps)
+    with torch.no_grad():
+        x_tst=phones.to(device).unsqueeze(0)
+        tones=tones.to(device).unsqueeze(0)
+        lang_ids=lang_ids.to(device).unsqueeze(0)
+        bert = bert.to(device).unsqueeze(0)
+        x_tst_lengths = torch.LongTensor([phones.size(0)]).to(device)
+        del phones
+        speakers = torch.LongTensor([hps.data.spk2id[sid]]).to(device)
+        audio = net_g.infer(x_tst, x_tst_lengths, speakers, tones, lang_ids, bert, sdp_ratio=sdp_ratio
+                           , noise_scale=noise_scale, noise_scale_w=noise_scale_w, length_scale=length_scale)[0][0,0].data.cpu().float().numpy()
+        del x_tst, tones, lang_ids, bert, x_tst_lengths, speakers
+        return audio
+
+def tts_fn(text, speaker, sdp_ratio, noise_scale, noise_scale_w, length_scale):
+    with torch.no_grad():
+        audio = infer(text, sdp_ratio=sdp_ratio, noise_scale=noise_scale, noise_scale_w=noise_scale_w, length_scale=length_scale, sid=speaker)
+    return "Success", (hps.data.sampling_rate, audio)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_dir", default="./logs/OUTPUT_MODEL/G_100.pth", help="path of your model")
+    parser.add_argument("--config_dir", default="./configs/config.json", help="path of your config file")
+    parser.add_argument("--share", default=False, help="make link public")
+    parser.add_argument("-d", "--debug", action="store_true", help="enable DEBUG-LEVEL log")
+
+    args = parser.parse_args()
+    if args.debug:
+        logger.info("Enable DEBUG-LEVEL log")
+        logging.basicConfig(level=logging.DEBUG)
+    hps = utils.get_hparams_from_file(args.config_dir)
+    device = "cuda:0" if torch.cuda.is_available() else "cpu"
+    '''
+    device = (
+        "cuda:0"
+        if torch.cuda.is_available()
+        else (
+            "mps"
+            if sys.platform == "darwin" and torch.backends.mps.is_available()
+            else "cpu"
+        )
+    )
+    '''
+    net_g = SynthesizerTrn(
+        len(symbols),
+        hps.data.filter_length // 2 + 1,
+        hps.train.segment_size // hps.data.hop_length,
+        n_speakers=hps.data.n_speakers,
+        **hps.model).to(device)
+    _ = net_g.eval()
+
+    _ = utils.load_checkpoint(args.model_dir, net_g, None, skip_optimizer=True)
+
+    speaker_ids = hps.data.spk2id
+    speakers = list(speaker_ids.keys())
+    with gr.Blocks() as app:
+        with gr.Row():
+            with gr.Column():
+                text = gr.TextArea(label="Text", placeholder="Input Text Here",
+                                      value="那是流萤吗，是明灭迷离，天真绮丽的憧憬。那是尘埃吧，是虚无纷飞，终将落地的谎言。")
+                speaker = gr.Dropdown(choices=speakers, value=speakers[0], label='Speaker')
+                sdp_ratio = gr.Slider(minimum=0, maximum=1, value=0.2, step=0.1, label='语调变化')
+                noise_scale = gr.Slider(minimum=0.1, maximum=1.5, value=0.6, step=0.1, label='感情变化')
+                noise_scale_w = gr.Slider(minimum=0.1, maximum=1.4, value=0.8, step=0.1, label='音节发音长度变化')
+                length_scale = gr.Slider(minimum=0.1, maximum=2, value=1, step=0.1, label='语速')
+                btn = gr.Button("开启AI语音之旅吧!", variant="primary")
+            with gr.Column():
+                text_output = gr.Textbox(label="Message")
+                audio_output = gr.Audio(label="Output Audio")
+
+        btn.click(tts_fn,
+                inputs=[text, speaker, sdp_ratio, noise_scale, noise_scale_w, length_scale],
+                outputs=[text_output, audio_output])
+    
+    webbrowser.open("http://127.0.0.1:6006")
+    app.launch(server_port=6006, show_error=True)

inference_webui.py.old

@@ -0,0 +1,142 @@
+import os
+import numpy as np
+import torch
+from torch import no_grad, LongTensor
+import argparse
+import commons
+from mel_processing import spectrogram_torch
+import utils
+from models import SynthesizerTrn
+import gradio as gr
+import librosa
+import webbrowser
+import time
+import commons
+import utils
+from models import SynthesizerTrn
+from text.symbols import symbols
+from text import cleaned_text_to_sequence,_symbol_to_id, get_bert
+from text.cleaner import clean_text
+from scipy.io import wavfile
+
+
+device = "cuda:0" if torch.cuda.is_available() else "cpu"
+import logging
+logging.getLogger("PIL").setLevel(logging.WARNING)
+logging.getLogger("urllib3").setLevel(logging.WARNING)
+logging.getLogger("markdown_it").setLevel(logging.WARNING)
+logging.getLogger("httpx").setLevel(logging.WARNING)
+logging.getLogger("asyncio").setLevel(logging.WARNING)
+
+language_marks = {
+    "简体中文": "[ZH]",
+}
+lang = ['简体中文']
+def get_text(text, language_str, hps):
+    norm_text, phone, tone, word2ph = clean_text(text, language_str)
+    print([f"{p}{t}" for p, t in zip(phone, tone)])
+    phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str)
+
+    if hps.data.add_blank:
+        phone = commons.intersperse(phone, 0)
+        tone = commons.intersperse(tone, 0)
+        language = commons.intersperse(language, 0)
+        for i in range(len(word2ph)):
+            word2ph[i] = word2ph[i] * 2
+        word2ph[0] += 1
+    bert = get_bert(norm_text, word2ph, language_str)
+
+    assert bert.shape[-1] == len(phone)
+
+    phone = torch.LongTensor(phone)
+    tone = torch.LongTensor(tone)
+    language = torch.LongTensor(language)
+
+    return bert, phone, tone, language
+'''
+def create_tts_fn(model, hps, speaker_ids):
+    def tts_fn(text, speaker, language, speed):
+        if language is not None:
+            text = language_marks[language] + text + language_marks[language]
+        speaker_id = speaker_ids[speaker]
+        stn_tst = get_text(text, hps, False)
+        with no_grad():
+            x_tst = stn_tst.unsqueeze(0).to(device)
+            x_tst_lengths = LongTensor([stn_tst.size(0)]).to(device)
+            sid = LongTensor([speaker_id]).to(device)
+            audio = model.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=.667, noise_scale_w=0.8,
+                                length_scale=1.0 / speed)[0][0, 0].data.cpu().float().numpy()
+        del stn_tst, x_tst, x_tst_lengths, sid
+        return "Success", (hps.data.sampling_rate, audio)
+
+    return tts_fn
+'''
+dev='cuda'
+def infer(text, sdp_ratio, noise_scale, noise_scale_w,length_scale,sid):
+    bert, phones, tones, lang_ids = get_text(text,"ZH", hps,)
+    print(sid)
+    with torch.no_grad():
+        x_tst=phones.to(dev).unsqueeze(0)
+        tones=tones.to(dev).unsqueeze(0)
+        lang_ids=lang_ids.to(dev).unsqueeze(0)
+        bert = bert.to(dev).unsqueeze(0)
+        x_tst_lengths = torch.LongTensor([phones.size(0)]).to(dev)
+        del phones
+        speakers = torch.LongTensor([hps.data.spk2id[sid]]).to(dev)
+        audio = net_g.infer(x_tst, x_tst_lengths, speakers, tones, lang_ids,bert, sdp_ratio=sdp_ratio
+                           , noise_scale=noise_scale, noise_scale_w=noise_scale_w, length_scale=length_scale)[0][0,0].data.cpu().float().numpy()
+        del x_tst, tones, lang_ids, bert, x_tst_lengths, speakers
+        return "Success",(hps.data.sampling_rate, audio)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_dir", default="./G_latest.pth", help="directory to your fine-tuned model")
+    parser.add_argument("--config_dir", default="./configs\config.json", help="directory to your model config file")
+    parser.add_argument("--share", default=False, help="make link public (used in colab)")
+
+    args = parser.parse_args()
+    hps = utils.get_hparams_from_file(args.config_dir)
+
+
+    net_g = SynthesizerTrn(
+    len(symbols),
+    hps.data.filter_length // 2 + 1,
+    hps.train.segment_size // hps.data.hop_length,
+    n_speakers=hps.data.n_speakers,
+    **hps.model).to(dev)
+    _ = net_g.eval()
+
+    _ = utils.load_checkpoint(args.model_dir, net_g, None,skip_optimizer=True)
+
+    speaker_ids = hps.data.spk2id
+    speakers = list(hps.data.spk2id.keys())
+    #inf = infer(net_g, hps, speaker_ids)
+    app = gr.Blocks()
+    with app:
+        with gr.Tab("Text-to-Speech"):
+            with gr.Row():
+                with gr.Column():
+                    textbox = gr.TextArea(label="Text",
+                                          placeholder="Type your sentence here",
+                                          value="生活就像海洋，只有意志坚强的人，才能到达彼岸。", elem_id=f"tts-input")
+                    # select character
+                    char_dropdown = gr.Dropdown(choices=speakers, value=speakers[0], label='character')
+                    language_dropdown = gr.Dropdown(choices=lang, value=lang[0], label='language')
+                    sdp_ratio = gr.Slider(minimum=0.1, maximum=0.9, value=0.2, step=0.1,
+                                                label='SDP/DP混合比-语调方差')
+                    noise_scale = gr.Slider(minimum=0.1, maximum=1.5, value=0.5, step=0.1,
+                                                label='noise/感情变化')
+                    noise_scale_w = gr.Slider(minimum=0.1, maximum=1.4, value=0.9, step=0.1,
+                                                label='noisew/音节发音长度变化')
+                    length_scale = gr.Slider(minimum=0.1, maximum=2, value=1.0, step=0.1,
+                                                label='length/语速')
+                with gr.Column():
+                    text_output = gr.Textbox(label="Message")
+                    audio_output = gr.Audio(label="Output Audio", elem_id="tts-audio")
+                    btn = gr.Button("Generate!")
+                    btn.click(infer,
+                              inputs=[textbox,sdp_ratio,noise_scale,noise_scale_w,length_scale,char_dropdown],
+                              outputs=[text_output, audio_output])
+    webbrowser.open("http://127.0.0.1:7860")
+    app.launch(share=args.share)
+

losses.py

@@ -0,0 +1,61 @@
+import torch 
+from torch.nn import functional as F
+
+import commons
+
+
+def feature_loss(fmap_r, fmap_g):
+  loss = 0
+  for dr, dg in zip(fmap_r, fmap_g):
+    for rl, gl in zip(dr, dg):
+      rl = rl.float().detach()
+      gl = gl.float()
+      loss += torch.mean(torch.abs(rl - gl))
+
+  return loss * 2 
+
+
+def discriminator_loss(disc_real_outputs, disc_generated_outputs):
+  loss = 0
+  r_losses = []
+  g_losses = []
+  for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
+    dr = dr.float()
+    dg = dg.float()
+    r_loss = torch.mean((1-dr)**2)
+    g_loss = torch.mean(dg**2)
+    loss += (r_loss + g_loss)
+    r_losses.append(r_loss.item())
+    g_losses.append(g_loss.item())
+
+  return loss, r_losses, g_losses
+
+
+def generator_loss(disc_outputs):
+  loss = 0
+  gen_losses = []
+  for dg in disc_outputs:
+    dg = dg.float()
+    l = torch.mean((1-dg)**2)
+    gen_losses.append(l)
+    loss += l
+
+  return loss, gen_losses
+
+
+def kl_loss(z_p, logs_q, m_p, logs_p, z_mask):
+  """
+  z_p, logs_q: [b, h, t_t]
+  m_p, logs_p: [b, h, t_t]
+  """
+  z_p = z_p.float()
+  logs_q = logs_q.float()
+  m_p = m_p.float()
+  logs_p = logs_p.float()
+  z_mask = z_mask.float()
+
+  kl = logs_p - logs_q - 0.5
+  kl += 0.5 * ((z_p - m_p)**2) * torch.exp(-2. * logs_p)
+  kl = torch.sum(kl * z_mask)
+  l = kl / torch.sum(z_mask)
+  return l

mel_processing.py

@@ -0,0 +1,112 @@
+import math
+import os
+import random
+import torch
+from torch import nn
+import torch.nn.functional as F
+import torch.utils.data
+import numpy as np
+import librosa
+import librosa.util as librosa_util
+from librosa.util import normalize, pad_center, tiny
+from scipy.signal import get_window
+from scipy.io.wavfile import read
+from librosa.filters import mel as librosa_mel_fn
+
+MAX_WAV_VALUE = 32768.0
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    """
+    PARAMS
+    ------
+    C: compression factor
+    """
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    """
+    PARAMS
+    ------
+    C: compression factor used to compress
+    """
+    return torch.exp(x) / C
+
+
+def spectral_normalize_torch(magnitudes):
+    output = dynamic_range_compression_torch(magnitudes)
+    return output
+
+
+def spectral_de_normalize_torch(magnitudes):
+    output = dynamic_range_decompression_torch(magnitudes)
+    return output
+
+
+mel_basis = {}
+hann_window = {}
+
+
+def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
+    if torch.min(y) < -1.:
+        print('min value is ', torch.min(y))
+    if torch.max(y) > 1.:
+        print('max value is ', torch.max(y))
+
+    global hann_window
+    dtype_device = str(y.dtype) + '_' + str(y.device)
+    wnsize_dtype_device = str(win_size) + '_' + dtype_device
+    if wnsize_dtype_device not in hann_window:
+        hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
+
+    y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
+    y = y.squeeze(1)
+
+    spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
+                      center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
+
+    spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
+    return spec
+
+
+def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
+    global mel_basis
+    dtype_device = str(spec.dtype) + '_' + str(spec.device)
+    fmax_dtype_device = str(fmax) + '_' + dtype_device
+    if fmax_dtype_device not in mel_basis:
+        mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
+        mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device)
+    spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
+    spec = spectral_normalize_torch(spec)
+    return spec
+
+
+def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
+    if torch.min(y) < -1.:
+        print('min value is ', torch.min(y))
+    if torch.max(y) > 1.:
+        print('max value is ', torch.max(y))
+
+    global mel_basis, hann_window
+    dtype_device = str(y.dtype) + '_' + str(y.device)
+    fmax_dtype_device = str(fmax) + '_' + dtype_device
+    wnsize_dtype_device = str(win_size) + '_' + dtype_device
+    if fmax_dtype_device not in mel_basis:
+        mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
+        mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device)
+    if wnsize_dtype_device not in hann_window:
+        hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
+
+    y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
+    y = y.squeeze(1)
+
+    spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
+                          center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
+
+    spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
+
+    spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
+    spec = spectral_normalize_torch(spec)
+
+    return spec

models.py

@@ -0,0 +1,707 @@
+import copy
+import math
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+import commons
+import modules
+import attentions
+import monotonic_align
+
+from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
+from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
+
+from commons import init_weights, get_padding
+from text import symbols, num_tones, num_languages
+class DurationDiscriminator(nn.Module): #vits2
+  def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):
+    super().__init__()
+
+    self.in_channels = in_channels
+    self.filter_channels = filter_channels
+    self.kernel_size = kernel_size
+    self.p_dropout = p_dropout
+    self.gin_channels = gin_channels
+
+    self.drop = nn.Dropout(p_dropout)
+    self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size//2)
+    self.norm_1 = modules.LayerNorm(filter_channels)
+    self.conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2)
+    self.norm_2 = modules.LayerNorm(filter_channels)
+    self.dur_proj = nn.Conv1d(1, filter_channels, 1)
+
+    self.pre_out_conv_1 = nn.Conv1d(2*filter_channels, filter_channels, kernel_size, padding=kernel_size//2)
+    self.pre_out_norm_1 = modules.LayerNorm(filter_channels)
+    self.pre_out_conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2)
+    self.pre_out_norm_2 = modules.LayerNorm(filter_channels)
+
+    if gin_channels != 0:
+       self.cond = nn.Conv1d(gin_channels, in_channels, 1)
+
+    self.output_layer = nn.Sequential(
+        nn.Linear(filter_channels, 1), 
+        nn.Sigmoid() 
+    )
+
+  def forward_probability(self, x, x_mask, dur, g=None):
+    dur = self.dur_proj(dur)
+    x = torch.cat([x, dur], dim=1)
+    x = self.pre_out_conv_1(x * x_mask)
+    x = torch.relu(x)
+    x = self.pre_out_norm_1(x)
+    x = self.drop(x)
+    x = self.pre_out_conv_2(x * x_mask)
+    x = torch.relu(x)
+    x = self.pre_out_norm_2(x)
+    x = self.drop(x)
+    x = x * x_mask
+    x = x.transpose(1, 2)
+    output_prob = self.output_layer(x)
+    return output_prob
+
+  def forward(self, x, x_mask, dur_r, dur_hat, g=None):
+    x = torch.detach(x)
+    if g is not None:
+       g = torch.detach(g)
+       x = x + self.cond(g)
+    x = self.conv_1(x * x_mask)
+    x = torch.relu(x)
+    x = self.norm_1(x)
+    x = self.drop(x)
+    x = self.conv_2(x * x_mask)
+    x = torch.relu(x)
+    x = self.norm_2(x)
+    x = self.drop(x)
+
+    output_probs = []
+    for dur in [dur_r, dur_hat]:
+      output_prob = self.forward_probability(x, x_mask, dur, g)
+      output_probs.append(output_prob)
+
+    return output_probs
+
+class TransformerCouplingBlock(nn.Module):
+    def __init__(self,
+                 channels,
+                 hidden_channels,
+                 filter_channels,
+                 n_heads,
+                 n_layers,
+                 kernel_size,
+                 p_dropout,
+                 n_flows=4,
+                 gin_channels=0,
+                 share_parameter=False
+                 ):
+
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+
+        self.wn = attentions.FFT(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, isflow = True, gin_channels = self.gin_channels) if share_parameter else None
+
+        for i in range(n_flows):
+            self.flows.append(
+                modules.TransformerCouplingLayer(channels, hidden_channels, kernel_size, n_layers, n_heads, p_dropout, filter_channels, mean_only=True, wn_sharing_parameter=self.wn, gin_channels = self.gin_channels))
+            self.flows.append(modules.Flip())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+class StochasticDurationPredictor(nn.Module):
+    def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):
+        super().__init__()
+        filter_channels = in_channels  # it needs to be removed from future version.
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.log_flow = modules.Log()
+        self.flows = nn.ModuleList()
+        self.flows.append(modules.ElementwiseAffine(2))
+        for i in range(n_flows):
+            self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
+            self.flows.append(modules.Flip())
+
+        self.post_pre = nn.Conv1d(1, filter_channels, 1)
+        self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
+        self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
+        self.post_flows = nn.ModuleList()
+        self.post_flows.append(modules.ElementwiseAffine(2))
+        for i in range(4):
+            self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
+            self.post_flows.append(modules.Flip())
+
+        self.pre = nn.Conv1d(in_channels, filter_channels, 1)
+        self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
+        self.convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
+
+    def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
+        x = torch.detach(x)
+        x = self.pre(x)
+        if g is not None:
+            g = torch.detach(g)
+            x = x + self.cond(g)
+        x = self.convs(x, x_mask)
+        x = self.proj(x) * x_mask
+
+        if not reverse:
+            flows = self.flows
+            assert w is not None
+
+            logdet_tot_q = 0
+            h_w = self.post_pre(w)
+            h_w = self.post_convs(h_w, x_mask)
+            h_w = self.post_proj(h_w) * x_mask
+            e_q = torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype) * x_mask
+            z_q = e_q
+            for flow in self.post_flows:
+                z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
+                logdet_tot_q += logdet_q
+            z_u, z1 = torch.split(z_q, [1, 1], 1)
+            u = torch.sigmoid(z_u) * x_mask
+            z0 = (w - u) * x_mask
+            logdet_tot_q += torch.sum((F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1, 2])
+            logq = torch.sum(-0.5 * (math.log(2 * math.pi) + (e_q ** 2)) * x_mask, [1, 2]) - logdet_tot_q
+
+            logdet_tot = 0
+            z0, logdet = self.log_flow(z0, x_mask)
+            logdet_tot += logdet
+            z = torch.cat([z0, z1], 1)
+            for flow in flows:
+                z, logdet = flow(z, x_mask, g=x, reverse=reverse)
+                logdet_tot = logdet_tot + logdet
+            nll = torch.sum(0.5 * (math.log(2 * math.pi) + (z ** 2)) * x_mask, [1, 2]) - logdet_tot
+            return nll + logq  # [b]
+        else:
+            flows = list(reversed(self.flows))
+            flows = flows[:-2] + [flows[-1]]  # remove a useless vflow
+            z = torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype) * noise_scale
+            for flow in flows:
+                z = flow(z, x_mask, g=x, reverse=reverse)
+            z0, z1 = torch.split(z, [1, 1], 1)
+            logw = z0
+            return logw
+
+
+class DurationPredictor(nn.Module):
+    def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.gin_channels = gin_channels
+
+        self.drop = nn.Dropout(p_dropout)
+        self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size // 2)
+        self.norm_1 = modules.LayerNorm(filter_channels)
+        self.conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size // 2)
+        self.norm_2 = modules.LayerNorm(filter_channels)
+        self.proj = nn.Conv1d(filter_channels, 1, 1)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, in_channels, 1)
+
+    def forward(self, x, x_mask, g=None):
+        x = torch.detach(x)
+        if g is not None:
+            g = torch.detach(g)
+            x = x + self.cond(g)
+        x = self.conv_1(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_1(x)
+        x = self.drop(x)
+        x = self.conv_2(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_2(x)
+        x = self.drop(x)
+        x = self.proj(x * x_mask)
+        return x * x_mask
+
+
+class TextEncoder(nn.Module):
+    def __init__(self,
+                 n_vocab,
+                 out_channels,
+                 hidden_channels,
+                 filter_channels,
+                 n_heads,
+                 n_layers,
+                 kernel_size,
+                 p_dropout,
+                 gin_channels=0):
+        super().__init__()
+        self.n_vocab = n_vocab
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.gin_channels = gin_channels
+        self.emb = nn.Embedding(len(symbols), hidden_channels)
+        nn.init.normal_(self.emb.weight, 0.0, hidden_channels ** -0.5)
+        self.tone_emb = nn.Embedding(num_tones, hidden_channels)
+        nn.init.normal_(self.tone_emb.weight, 0.0, hidden_channels ** -0.5)
+        self.language_emb = nn.Embedding(num_languages, hidden_channels)
+        nn.init.normal_(self.language_emb.weight, 0.0, hidden_channels ** -0.5)
+        self.bert_proj = nn.Conv1d(1024, hidden_channels, 1)
+
+        self.encoder = attentions.Encoder(
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            gin_channels=self.gin_channels)
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, tone, language, bert, g=None):
+        x = (self.emb(x)+ self.tone_emb(tone)+ self.language_emb(language)+self.bert_proj(bert).transpose(1,2)) * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
+
+        x = self.encoder(x * x_mask, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return x, m, logs, x_mask
+
+
+class ResidualCouplingBlock(nn.Module):
+    def __init__(self,
+                 channels,
+                 hidden_channels,
+                 kernel_size,
+                 dilation_rate,
+                 n_layers,
+                 n_flows=4,
+                 gin_channels=0):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers,
+                                              gin_channels=gin_channels, mean_only=True))
+            self.flows.append(modules.Flip())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+
+class PosteriorEncoder(nn.Module):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 hidden_channels,
+                 kernel_size,
+                 dilation_rate,
+                 n_layers,
+                 gin_channels=0):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+
+        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
+        self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, g=None):
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
+        x = self.pre(x) * x_mask
+        x = self.enc(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
+        return z, m, logs, x_mask
+
+
+class Generator(torch.nn.Module):
+    def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates,
+                 upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):
+        super(Generator, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
+        resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(weight_norm(
+                ConvTranspose1d(upsample_initial_channel // (2 ** i), upsample_initial_channel // (2 ** (i + 1)),
+                                k, u, padding=(k - u) // 2)))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+    def forward(self, x, g=None):
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        print('Removing weight norm...')
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+class DiscriminatorP(torch.nn.Module):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        self.use_spectral_norm = use_spectral_norm
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList([
+            norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
+            norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
+            norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
+            norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
+            norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
+        ])
+        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class DiscriminatorS(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList([
+            norm_f(Conv1d(1, 16, 15, 1, padding=7)),
+            norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
+            norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
+            norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
+            norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
+            norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
+        ])
+        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        fmap = []
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminator, self).__init__()
+        periods = [2, 3, 5, 7, 11]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
+        self.discriminators = nn.ModuleList(discs)
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+class ReferenceEncoder(nn.Module):
+    '''
+    inputs --- [N, Ty/r, n_mels*r]  mels
+    outputs --- [N, ref_enc_gru_size]
+    '''
+
+    def __init__(self, spec_channels, gin_channels=0):
+
+        super().__init__()
+        self.spec_channels = spec_channels
+        ref_enc_filters = [32, 32, 64, 64, 128, 128]
+        K = len(ref_enc_filters)
+        filters = [1] + ref_enc_filters
+        convs = [weight_norm(nn.Conv2d(in_channels=filters[i],
+                                       out_channels=filters[i + 1],
+                                       kernel_size=(3, 3),
+                                       stride=(2, 2),
+                                       padding=(1, 1))) for i in range(K)]
+        self.convs = nn.ModuleList(convs)
+        # self.wns = nn.ModuleList([weight_norm(num_features=ref_enc_filters[i]) for i in range(K)])
+
+        out_channels = self.calculate_channels(spec_channels, 3, 2, 1, K)
+        self.gru = nn.GRU(input_size=ref_enc_filters[-1] * out_channels,
+                          hidden_size=256 // 2,
+                          batch_first=True)
+        self.proj = nn.Linear(128, gin_channels)
+
+    def forward(self, inputs, mask=None):
+        N = inputs.size(0)
+        out = inputs.view(N, 1, -1, self.spec_channels)  # [N, 1, Ty, n_freqs]
+        for conv in self.convs:
+            out = conv(out)
+            # out = wn(out)
+            out = F.relu(out)  # [N, 128, Ty//2^K, n_mels//2^K]
+
+        out = out.transpose(1, 2)  # [N, Ty//2^K, 128, n_mels//2^K]
+        T = out.size(1)
+        N = out.size(0)
+        out = out.contiguous().view(N, T, -1)  # [N, Ty//2^K, 128*n_mels//2^K]
+
+        self.gru.flatten_parameters()
+        memory, out = self.gru(out)  # out --- [1, N, 128]
+
+        return self.proj(out.squeeze(0))
+
+    def calculate_channels(self, L, kernel_size, stride, pad, n_convs):
+        for i in range(n_convs):
+            L = (L - kernel_size + 2 * pad) // stride + 1
+        return L
+
+
+class SynthesizerTrn(nn.Module):
+    """
+  Synthesizer for Training
+  """
+
+    def __init__(self,
+                 n_vocab,
+                 spec_channels,
+                 segment_size,
+                 inter_channels,
+                 hidden_channels,
+                 filter_channels,
+                 n_heads,
+                 n_layers,
+                 kernel_size,
+                 p_dropout,
+                 resblock,
+                 resblock_kernel_sizes,
+                 resblock_dilation_sizes,
+                 upsample_rates,
+                 upsample_initial_channel,
+                 upsample_kernel_sizes,
+                 n_speakers=256,
+                 gin_channels=256,
+                 use_sdp=True,
+                 n_flow_layer = 4,
+                 n_layers_trans_flow = 3,
+                 flow_share_parameter = False,
+                 use_transformer_flow = True,
+                 **kwargs):
+
+        super().__init__()
+        self.n_vocab = n_vocab
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.n_speakers = n_speakers
+        self.gin_channels = gin_channels
+        self.n_layers_trans_flow = n_layers_trans_flow
+        self.use_spk_conditioned_encoder = kwargs.get("use_spk_conditioned_encoder", True)
+        self.use_sdp = use_sdp
+        self.use_noise_scaled_mas = kwargs.get("use_noise_scaled_mas", False)
+        self.mas_noise_scale_initial = kwargs.get("mas_noise_scale_initial", 0.01)
+        self.noise_scale_delta = kwargs.get("noise_scale_delta", 2e-6)
+        self.current_mas_noise_scale = self.mas_noise_scale_initial
+        if self.use_spk_conditioned_encoder and gin_channels > 0:
+            self.enc_gin_channels = gin_channels
+        self.enc_p = TextEncoder(n_vocab,
+                                 inter_channels,
+                                 hidden_channels,
+                                 filter_channels,
+                                 n_heads,
+                                 n_layers,
+                                 kernel_size,
+                                 p_dropout,
+                                 gin_channels=self.enc_gin_channels)
+        self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates,
+                             upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
+        self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16,
+                                      gin_channels=gin_channels)
+        if use_transformer_flow:
+            self.flow = TransformerCouplingBlock(inter_channels, hidden_channels, filter_channels, n_heads, n_layers_trans_flow, 5, p_dropout, n_flow_layer, gin_channels=gin_channels,share_parameter= flow_share_parameter)
+        else:
+            self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, n_flow_layer, gin_channels=gin_channels)
+        self.sdp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)
+        self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)
+        
+        if n_speakers >= 1:
+            self.emb_g = nn.Embedding(n_speakers, gin_channels)
+        else:
+            self.ref_enc = ReferenceEncoder(spec_channels, gin_channels)
+
+    def forward(self, x, x_lengths, y, y_lengths, sid, tone, language, bert):
+        if self.n_speakers > 0:
+            g = self.emb_g(sid).unsqueeze(-1)  # [b, h, 1]
+        else:
+            g = self.ref_enc(y.transpose(1,2)).unsqueeze(-1)
+        x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths, tone, language, bert,g=g)
+        z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
+        z_p = self.flow(z, y_mask, g=g)
+
+        with torch.no_grad():
+            # negative cross-entropy
+            s_p_sq_r = torch.exp(-2 * logs_p)  # [b, d, t]
+            neg_cent1 = torch.sum(-0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True)  # [b, 1, t_s]
+            neg_cent2 = torch.matmul(-0.5 * (z_p ** 2).transpose(1, 2),
+                                     s_p_sq_r)  # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]
+            neg_cent3 = torch.matmul(z_p.transpose(1, 2), (m_p * s_p_sq_r))  # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]
+            neg_cent4 = torch.sum(-0.5 * (m_p ** 2) * s_p_sq_r, [1], keepdim=True)  # [b, 1, t_s]
+            neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4
+            if self.use_noise_scaled_mas:
+               epsilon = torch.std(neg_cent) * torch.randn_like(neg_cent) * self.current_mas_noise_scale
+               neg_cent = neg_cent + epsilon
+
+            attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
+            attn = monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1)).unsqueeze(1).detach()
+
+        w = attn.sum(2)
+
+        l_length_sdp = self.sdp(x, x_mask, w, g=g)
+        l_length_sdp = l_length_sdp / torch.sum(x_mask)
+        
+        logw_ = torch.log(w + 1e-6) * x_mask
+        logw = self.dp(x, x_mask, g=g)
+        l_length_dp = torch.sum((logw - logw_) ** 2, [1, 2]) / torch.sum(x_mask)  # for averaging
+
+        l_length = l_length_dp + l_length_sdp
+
+        # expand prior
+        m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)
+        logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)
+
+        z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
+        o = self.dec(z_slice, g=g)
+        return o, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q), (x, logw, logw_)
+      
+    def infer(self, x, x_lengths, sid, tone, language, bert, noise_scale=.667, length_scale=1, noise_scale_w=0.8, max_len=None, sdp_ratio=0,y=None):
+        #x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths, tone, language, bert)
+        # g = self.gst(y)
+        if self.n_speakers > 0:
+            g = self.emb_g(sid).unsqueeze(-1)  # [b, h, 1]
+        else:
+            g = self.ref_enc(y.transpose(1,2)).unsqueeze(-1)
+        x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths, tone, language, bert,g=g)
+        logw = self.sdp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w) * (sdp_ratio) + self.dp(x, x_mask, g=g) * (1 - sdp_ratio)
+        w = torch.exp(logw) * x_mask * length_scale
+        w_ceil = torch.ceil(w)
+        y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
+        y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
+        attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
+        attn = commons.generate_path(w_ceil, attn_mask)
+
+        m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)  # [b, t', t], [b, t, d] -> [b, d, t']
+        logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1,
+                                                                                 2)  # [b, t', t], [b, t, d] -> [b, d, t']
+
+        z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
+        z = self.flow(z_p, y_mask, g=g, reverse=True)
+        o = self.dec((z * y_mask)[:, :, :max_len], g=g)
+        return o, attn, y_mask, (z, z_p, m_p, logs_p)

modules.py

@@ -0,0 +1,452 @@
+import copy
+import math
+import numpy as np
+import scipy
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
+from torch.nn.utils import weight_norm, remove_weight_norm
+
+import commons
+from commons import init_weights, get_padding
+from transforms import piecewise_rational_quadratic_transform
+from attentions import Encoder
+
+LRELU_SLOPE = 0.1
+
+class LayerNorm(nn.Module):
+  def __init__(self, channels, eps=1e-5):
+    super().__init__()
+    self.channels = channels
+    self.eps = eps
+
+    self.gamma = nn.Parameter(torch.ones(channels))
+    self.beta = nn.Parameter(torch.zeros(channels))
+
+  def forward(self, x):
+    x = x.transpose(1, -1)
+    x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+    return x.transpose(1, -1)
+    
+class ConvReluNorm(nn.Module):
+  def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
+    super().__init__()
+    self.in_channels = in_channels
+    self.hidden_channels = hidden_channels
+    self.out_channels = out_channels
+    self.kernel_size = kernel_size
+    self.n_layers = n_layers
+    self.p_dropout = p_dropout
+    assert n_layers > 1, "Number of layers should be larger than 0."
+
+    self.conv_layers = nn.ModuleList()
+    self.norm_layers = nn.ModuleList()
+    self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size//2))
+    self.norm_layers.append(LayerNorm(hidden_channels))
+    self.relu_drop = nn.Sequential(
+        nn.ReLU(),
+        nn.Dropout(p_dropout))
+    for _ in range(n_layers-1):
+      self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size//2))
+      self.norm_layers.append(LayerNorm(hidden_channels))
+    self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
+    self.proj.weight.data.zero_()
+    self.proj.bias.data.zero_()
+
+  def forward(self, x, x_mask):
+    x_org = x
+    for i in range(self.n_layers):
+      x = self.conv_layers[i](x * x_mask)
+      x = self.norm_layers[i](x)
+      x = self.relu_drop(x)
+    x = x_org + self.proj(x)
+    return x * x_mask
+
+
+class DDSConv(nn.Module):
+  """
+  Dialted and Depth-Separable Convolution
+  """
+  def __init__(self, channels, kernel_size, n_layers, p_dropout=0.):
+    super().__init__()
+    self.channels = channels
+    self.kernel_size = kernel_size
+    self.n_layers = n_layers
+    self.p_dropout = p_dropout
+
+    self.drop = nn.Dropout(p_dropout)
+    self.convs_sep = nn.ModuleList()
+    self.convs_1x1 = nn.ModuleList()
+    self.norms_1 = nn.ModuleList()
+    self.norms_2 = nn.ModuleList()
+    for i in range(n_layers):
+      dilation = kernel_size ** i
+      padding = (kernel_size * dilation - dilation) // 2
+      self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size, 
+          groups=channels, dilation=dilation, padding=padding
+      ))
+      self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
+      self.norms_1.append(LayerNorm(channels))
+      self.norms_2.append(LayerNorm(channels))
+
+  def forward(self, x, x_mask, g=None):
+    if g is not None:
+      x = x + g
+    for i in range(self.n_layers):
+      y = self.convs_sep[i](x * x_mask)
+      y = self.norms_1[i](y)
+      y = F.gelu(y)
+      y = self.convs_1x1[i](y)
+      y = self.norms_2[i](y)
+      y = F.gelu(y)
+      y = self.drop(y)
+      x = x + y
+    return x * x_mask
+
+
+class WN(torch.nn.Module):
+  def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
+    super(WN, self).__init__()
+    assert(kernel_size % 2 == 1)
+    self.hidden_channels =hidden_channels
+    self.kernel_size = kernel_size,
+    self.dilation_rate = dilation_rate
+    self.n_layers = n_layers
+    self.gin_channels = gin_channels
+    self.p_dropout = p_dropout
+
+    self.in_layers = torch.nn.ModuleList()
+    self.res_skip_layers = torch.nn.ModuleList()
+    self.drop = nn.Dropout(p_dropout)
+
+    if gin_channels != 0:
+      cond_layer = torch.nn.Conv1d(gin_channels, 2*hidden_channels*n_layers, 1)
+      self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name='weight')
+
+    for i in range(n_layers):
+      dilation = dilation_rate ** i
+      padding = int((kernel_size * dilation - dilation) / 2)
+      in_layer = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, kernel_size,
+                                 dilation=dilation, padding=padding)
+      in_layer = torch.nn.utils.weight_norm(in_layer, name='weight')
+      self.in_layers.append(in_layer)
+
+      # last one is not necessary
+      if i < n_layers - 1:
+        res_skip_channels = 2 * hidden_channels
+      else:
+        res_skip_channels = hidden_channels
+
+      res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
+      res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name='weight')
+      self.res_skip_layers.append(res_skip_layer)
+
+  def forward(self, x, x_mask, g=None, **kwargs):
+    output = torch.zeros_like(x)
+    n_channels_tensor = torch.IntTensor([self.hidden_channels])
+
+    if g is not None:
+      g = self.cond_layer(g)
+
+    for i in range(self.n_layers):
+      x_in = self.in_layers[i](x)
+      if g is not None:
+        cond_offset = i * 2 * self.hidden_channels
+        g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]
+      else:
+        g_l = torch.zeros_like(x_in)
+
+      acts = commons.fused_add_tanh_sigmoid_multiply(
+          x_in,
+          g_l,
+          n_channels_tensor)
+      acts = self.drop(acts)
+
+      res_skip_acts = self.res_skip_layers[i](acts)
+      if i < self.n_layers - 1:
+        res_acts = res_skip_acts[:,:self.hidden_channels,:]
+        x = (x + res_acts) * x_mask
+        output = output + res_skip_acts[:,self.hidden_channels:,:]
+      else:
+        output = output + res_skip_acts
+    return output * x_mask
+
+  def remove_weight_norm(self):
+    if self.gin_channels != 0:
+      torch.nn.utils.remove_weight_norm(self.cond_layer)
+    for l in self.in_layers:
+      torch.nn.utils.remove_weight_norm(l)
+    for l in self.res_skip_layers:
+     torch.nn.utils.remove_weight_norm(l)
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.convs1 = nn.ModuleList([
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
+                               padding=get_padding(kernel_size, dilation[0]))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
+                               padding=get_padding(kernel_size, dilation[1]))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
+                               padding=get_padding(kernel_size, dilation[2])))
+        ])
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList([
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1)))
+        ])
+        self.convs2.apply(init_weights)
+
+    def forward(self, x, x_mask=None):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c2(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class ResBlock2(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        self.convs = nn.ModuleList([
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
+                               padding=get_padding(kernel_size, dilation[0]))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
+                               padding=get_padding(kernel_size, dilation[1])))
+        ])
+        self.convs.apply(init_weights)
+
+    def forward(self, x, x_mask=None):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+
+class Log(nn.Module):
+  def forward(self, x, x_mask, reverse=False, **kwargs):
+    if not reverse:
+      y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
+      logdet = torch.sum(-y, [1, 2])
+      return y, logdet
+    else:
+      x = torch.exp(x) * x_mask
+      return x
+    
+
+class Flip(nn.Module):
+  def forward(self, x, *args, reverse=False, **kwargs):
+    x = torch.flip(x, [1])
+    if not reverse:
+      logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
+      return x, logdet
+    else:
+      return x
+
+
+class ElementwiseAffine(nn.Module):
+  def __init__(self, channels):
+    super().__init__()
+    self.channels = channels
+    self.m = nn.Parameter(torch.zeros(channels,1))
+    self.logs = nn.Parameter(torch.zeros(channels,1))
+
+  def forward(self, x, x_mask, reverse=False, **kwargs):
+    if not reverse:
+      y = self.m + torch.exp(self.logs) * x
+      y = y * x_mask
+      logdet = torch.sum(self.logs * x_mask, [1,2])
+      return y, logdet
+    else:
+      x = (x - self.m) * torch.exp(-self.logs) * x_mask
+      return x
+
+
+class ResidualCouplingLayer(nn.Module):
+  def __init__(self,
+      channels,
+      hidden_channels,
+      kernel_size,
+      dilation_rate,
+      n_layers,
+      p_dropout=0,
+      gin_channels=0,
+      mean_only=False):
+    assert channels % 2 == 0, "channels should be divisible by 2"
+    super().__init__()
+    self.channels = channels
+    self.hidden_channels = hidden_channels
+    self.kernel_size = kernel_size
+    self.dilation_rate = dilation_rate
+    self.n_layers = n_layers
+    self.half_channels = channels // 2
+    self.mean_only = mean_only
+
+    self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+    self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
+    self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+    self.post.weight.data.zero_()
+    self.post.bias.data.zero_()
+
+  def forward(self, x, x_mask, g=None, reverse=False):
+    x0, x1 = torch.split(x, [self.half_channels]*2, 1)
+    h = self.pre(x0) * x_mask
+    h = self.enc(h, x_mask, g=g)
+    stats = self.post(h) * x_mask
+    if not self.mean_only:
+      m, logs = torch.split(stats, [self.half_channels]*2, 1)
+    else:
+      m = stats
+      logs = torch.zeros_like(m)
+
+    if not reverse:
+      x1 = m + x1 * torch.exp(logs) * x_mask
+      x = torch.cat([x0, x1], 1)
+      logdet = torch.sum(logs, [1,2])
+      return x, logdet
+    else:
+      x1 = (x1 - m) * torch.exp(-logs) * x_mask
+      x = torch.cat([x0, x1], 1)
+      return x
+
+
+class ConvFlow(nn.Module):
+  def __init__(self, in_channels, filter_channels, kernel_size, n_layers, num_bins=10, tail_bound=5.0):
+    super().__init__()
+    self.in_channels = in_channels
+    self.filter_channels = filter_channels
+    self.kernel_size = kernel_size
+    self.n_layers = n_layers
+    self.num_bins = num_bins
+    self.tail_bound = tail_bound
+    self.half_channels = in_channels // 2
+
+    self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
+    self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.)
+    self.proj = nn.Conv1d(filter_channels, self.half_channels * (num_bins * 3 - 1), 1)
+    self.proj.weight.data.zero_()
+    self.proj.bias.data.zero_()
+
+  def forward(self, x, x_mask, g=None, reverse=False):
+    x0, x1 = torch.split(x, [self.half_channels]*2, 1)
+    h = self.pre(x0)
+    h = self.convs(h, x_mask, g=g)
+    h = self.proj(h) * x_mask
+
+    b, c, t = x0.shape
+    h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2) # [b, cx?, t] -> [b, c, t, ?]
+
+    unnormalized_widths = h[..., :self.num_bins] / math.sqrt(self.filter_channels)
+    unnormalized_heights = h[..., self.num_bins:2*self.num_bins] / math.sqrt(self.filter_channels)
+    unnormalized_derivatives = h[..., 2 * self.num_bins:]
+
+    x1, logabsdet = piecewise_rational_quadratic_transform(x1,
+        unnormalized_widths,
+        unnormalized_heights,
+        unnormalized_derivatives,
+        inverse=reverse,
+        tails='linear',
+        tail_bound=self.tail_bound
+    )
+
+    x = torch.cat([x0, x1], 1) * x_mask
+    logdet = torch.sum(logabsdet * x_mask, [1,2])
+    if not reverse:
+        return x, logdet
+    else:
+        return x
+class TransformerCouplingLayer(nn.Module):
+  def __init__(self,
+      channels,
+      hidden_channels,
+      kernel_size,
+      n_layers,
+      n_heads,
+      p_dropout=0,
+      filter_channels=0,
+      mean_only=False,
+      wn_sharing_parameter=None,
+      gin_channels = 0
+      ):
+    assert channels % 2 == 0, "channels should be divisible by 2"
+    super().__init__()
+    self.channels = channels
+    self.hidden_channels = hidden_channels
+    self.kernel_size = kernel_size
+    self.n_layers = n_layers
+    self.half_channels = channels // 2
+    self.mean_only = mean_only
+
+    self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+    self.enc = Encoder(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, isflow = True, gin_channels = gin_channels) if wn_sharing_parameter is None else wn_sharing_parameter
+    self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+    self.post.weight.data.zero_()
+    self.post.bias.data.zero_()
+
+  def forward(self, x, x_mask, g=None, reverse=False):
+    x0, x1 = torch.split(x, [self.half_channels]*2, 1)
+    h = self.pre(x0) * x_mask
+    h = self.enc(h, x_mask, g=g)
+    stats = self.post(h) * x_mask
+    if not self.mean_only:
+      m, logs = torch.split(stats, [self.half_channels]*2, 1)
+    else:
+      m = stats
+      logs = torch.zeros_like(m)
+
+    if not reverse:
+      x1 = m + x1 * torch.exp(logs) * x_mask
+      x = torch.cat([x0, x1], 1)
+      logdet = torch.sum(logs, [1,2])
+      return x, logdet
+    else:
+      x1 = (x1 - m) * torch.exp(-logs) * x_mask
+      x = torch.cat([x0, x1], 1)
+      return x
+
+    x1, logabsdet = piecewise_rational_quadratic_transform(x1,
+        unnormalized_widths,
+        unnormalized_heights,
+        unnormalized_derivatives,
+        inverse=reverse,
+        tails='linear',
+        tail_bound=self.tail_bound
+    )
+
+    x = torch.cat([x0, x1], 1) * x_mask
+    logdet = torch.sum(logabsdet * x_mask, [1,2])
+    if not reverse:
+        return x, logdet
+    else:
+        return x

monotonic_align/__init__.py

@@ -0,0 +1,19 @@
+import numpy as np
+import torch
+from .monotonic_align.core import maximum_path_c
+
+
+def maximum_path(neg_cent, mask):
+  """ Cython optimized version.
+  neg_cent: [b, t_t, t_s]
+  mask: [b, t_t, t_s]
+  """
+  device = neg_cent.device
+  dtype = neg_cent.dtype
+  neg_cent = neg_cent.data.cpu().numpy().astype(np.float32)
+  path = np.zeros(neg_cent.shape, dtype=np.int32)
+
+  t_t_max = mask.sum(1)[:, 0].data.cpu().numpy().astype(np.int32)
+  t_s_max = mask.sum(2)[:, 0].data.cpu().numpy().astype(np.int32)
+  maximum_path_c(path, neg_cent, t_t_max, t_s_max)
+  return torch.from_numpy(path).to(device=device, dtype=dtype)

monotonic_align/core.pyx

@@ -0,0 +1,42 @@
+cimport cython
+from cython.parallel import prange
+
+
+@cython.boundscheck(False)
+@cython.wraparound(False)
+cdef void maximum_path_each(int[:,::1] path, float[:,::1] value, int t_y, int t_x, float max_neg_val=-1e9) nogil:
+  cdef int x
+  cdef int y
+  cdef float v_prev
+  cdef float v_cur
+  cdef float tmp
+  cdef int index = t_x - 1
+
+  for y in range(t_y):
+    for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
+      if x == y:
+        v_cur = max_neg_val
+      else:
+        v_cur = value[y-1, x]
+      if x == 0:
+        if y == 0:
+          v_prev = 0.
+        else:
+          v_prev = max_neg_val
+      else:
+        v_prev = value[y-1, x-1]
+      value[y, x] += max(v_prev, v_cur)
+
+  for y in range(t_y - 1, -1, -1):
+    path[y, index] = 1
+    if index != 0 and (index == y or value[y-1, index] < value[y-1, index-1]):
+      index = index - 1
+
+
+@cython.boundscheck(False)
+@cython.wraparound(False)
+cpdef void maximum_path_c(int[:,:,::1] paths, float[:,:,::1] values, int[::1] t_ys, int[::1] t_xs) nogil:
+  cdef int b = paths.shape[0]
+  cdef int i
+  for i in prange(b, nogil=True):
+    maximum_path_each(paths[i], values[i], t_ys[i], t_xs[i])

monotonic_align/setup.py

@@ -0,0 +1,9 @@
+from distutils.core import setup
+from Cython.Build import cythonize
+import numpy
+
+setup(
+  name = 'monotonic_align',
+  ext_modules = cythonize("core.pyx"),
+  include_dirs=[numpy.get_include()]
+)

preprocess_text.py

@@ -0,0 +1,69 @@
+import json
+from random import shuffle
+
+import tqdm
+from text.cleaner import clean_text
+from collections import defaultdict
+import shutil
+stage = [1,2,3]
+
+transcription_path = 'filelists/short_character_anno.list'
+train_path = 'filelists/train.list'
+val_path = 'filelists/val.list'
+config_path = "configs/config.json"
+val_per_spk = 4
+max_val_total = 8
+
+if 1 in stage:
+    with open( transcription_path+'.cleaned', 'w', encoding='utf-8') as f:
+        for line in tqdm.tqdm(open(transcription_path, encoding='utf-8').readlines()):
+            try:
+                utt, spk, language, text = line.strip().split('|')
+                #language = "ZH"
+                norm_text, phones, tones, word2ph = clean_text(text, language)
+                f.write('{}|{}|{}|{}|{}|{}|{}\n'.format(utt, spk, language, norm_text, ' '.join(phones),
+                                                     " ".join([str(i) for i in tones]),
+                                                     " ".join([str(i) for i in word2ph])))
+            except:
+                print("err!", utt)
+
+if 2 in stage:
+    spk_utt_map = defaultdict(list)
+    spk_id_map = {}
+    current_sid = 0
+
+    with open( transcription_path+'.cleaned', encoding='utf-8') as f:
+        for line in f.readlines():
+            utt, spk, language, text, phones, tones, word2ph = line.strip().split('|')
+            spk_utt_map[spk].append(line)
+            if spk not in spk_id_map.keys():
+                spk_id_map[spk] = current_sid
+                current_sid += 1
+    train_list = []
+    val_list = []
+    for spk, utts in spk_utt_map.items():
+          shuffle(utts)
+          val_list+=utts[:val_per_spk]
+          train_list+=utts[val_per_spk:]
+    if len(val_list) > max_val_total:
+          train_list+=val_list[max_val_total:]
+          val_list = val_list[:max_val_total]
+     
+    with open( train_path,"w", encoding='utf-8') as f:
+         for line in train_list:
+            f.write(line)
+
+    file_path = transcription_path+'.cleaned'
+    shutil.copy(file_path,'./filelists/train.list')
+    
+    with open(val_path, "w", encoding='utf-8') as f:
+        for line in val_list:
+            f.write(line)
+
+if 3 in stage:
+    assert 2 in stage
+    config = json.load(open(config_path))
+    config['data']["n_speakers"] = current_sid #
+    config["data"]['spk2id'] = spk_id_map
+    with open(config_path, 'w', encoding='utf-8') as f:
+        json.dump(config, f, indent=2, ensure_ascii=False)

requirements.txt

@@ -0,0 +1,26 @@
+librosa==0.9.1
+matplotlib
+numpy
+numba
+phonemizer
+scipy
+tensorboard
+torch
+torchaudio
+torchvision
+Unidecode
+amfm_decompy
+jieba
+transformers
+pypinyin
+cn2an
+gradio
+av
+mecab-python3
+loguru
+unidic-lite
+cmudict
+fugashi
+num2words
+Cython==0.29.21
+openai-whisper

setup_ffmpeg.py

@@ -0,0 +1,55 @@
+import os
+import sys
+import re
+from pathlib import Path
+import winreg
+
+def check_ffmpeg_path():
+    path_list = os.environ['Path'].split(';')
+    ffmpeg_found = False
+
+    for path in path_list:
+        if 'ffmpeg' in path.lower() and 'bin' in path.lower():
+            ffmpeg_found = True
+            print("FFmpeg already installed.")
+            break
+
+    return ffmpeg_found
+
+def add_ffmpeg_path_to_user_variable():
+    ffmpeg_bin_path = Path('.\\ffmpeg\\bin')
+    if ffmpeg_bin_path.is_dir():
+        abs_path = str(ffmpeg_bin_path.resolve())
+        
+        try:
+            key = winreg.OpenKey(
+                winreg.HKEY_CURRENT_USER,
+                r"Environment",
+                0,
+                winreg.KEY_READ | winreg.KEY_WRITE
+            )
+            
+            try:
+                current_path, _ = winreg.QueryValueEx(key, "Path")
+                if abs_path not in current_path:
+                    new_path = f"{current_path};{abs_path}"
+                    winreg.SetValueEx(key, "Path", 0, winreg.REG_EXPAND_SZ, new_path)
+                    print(f"Added FFmpeg path to user variable 'Path': {abs_path}")
+                else:
+                    print("FFmpeg path already exists in the user variable 'Path'.")
+            finally:
+                winreg.CloseKey(key)
+        except WindowsError:
+            print("Error: Unable to modify user variable 'Path'.")
+            sys.exit(1)
+
+    else:
+        print("Error: ffmpeg\\bin folder not found in the current path.")
+        sys.exit(1)
+
+def main():
+    if not check_ffmpeg_path():
+       add_ffmpeg_path_to_user_variable()
+            
+if __name__ == "__main__":
+    main()

short_audio_transcribe.py

@@ -0,0 +1,122 @@
+import whisper
+import os
+import json
+import torchaudio
+import argparse
+import torch
+
+lang2token = {
+            'zh': "[ZH]",
+            'ja': "[JA]",
+            "en": "[EN]",
+        }
+def transcribe_one(audio_path):
+    # load audio and pad/trim it to fit 30 seconds
+    audio = whisper.load_audio(audio_path)
+    audio = whisper.pad_or_trim(audio)
+
+    # make log-Mel spectrogram and move to the same device as the model
+    mel = whisper.log_mel_spectrogram(audio).to(model.device)
+
+    # detect the spoken language
+    _, probs = model.detect_language(mel)
+    print(f"Detected language: {max(probs, key=probs.get)}")
+    lang = max(probs, key=probs.get)
+    # decode the audio
+    options = whisper.DecodingOptions(beam_size=5)
+    result = whisper.decode(model, mel, options)
+
+    # print the recognized text
+    print(result.text)
+    return lang, result.text
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--languages", default="CJE")
+    parser.add_argument("--whisper_size", default="medium")
+    args = parser.parse_args()
+    if args.languages == "CJE":
+        lang2token = {
+            'zh': "[ZH]",
+            'ja': "[JA]",
+            "en": "[EN]",
+        }
+    elif args.languages == "CJ":
+        lang2token = {
+            'zh': "[ZH]",
+            'ja': "[JA]",
+        }
+    elif args.languages == "C":
+        lang2token = {
+            'zh': "[ZH]",
+        }
+    assert (torch.cuda.is_available()), "Please enable GPU in order to run Whisper!"
+    model = whisper.load_model(args.whisper_size)
+    parent_dir = "./custom_character_voice/"
+    speaker_names = list(os.walk(parent_dir))[0][1]
+    speaker_annos = []
+    total_files = sum([len(files) for r, d, files in os.walk(parent_dir)])
+    # resample audios
+    # 2023/4/21: Get the target sampling rate
+    with open("./configs/config.json", 'r', encoding='utf-8') as f:
+        hps = json.load(f)
+    target_sr = hps['data']['sampling_rate']
+    processed_files = 0
+    for speaker in speaker_names:
+        for i, wavfile in enumerate(list(os.walk(parent_dir + speaker))[0][2]):
+            # try to load file as audio
+            if wavfile.startswith("processed_"):
+                continue
+            try:
+                wav, sr = torchaudio.load(parent_dir + speaker + "/" + wavfile, frame_offset=0, num_frames=-1, normalize=True,
+                                          channels_first=True)
+                wav = wav.mean(dim=0).unsqueeze(0)
+                if sr != target_sr:
+                    wav = torchaudio.transforms.Resample(orig_freq=sr, new_freq=target_sr)(wav)
+                if wav.shape[1] / sr > 20:
+                    print(f"{wavfile} too long, ignoring\n")
+                save_path = parent_dir + speaker + "/" + f"processed_{i}.wav"
+                torchaudio.save(save_path, wav, target_sr, channels_first=True)
+                # transcribe text
+                lang, text = transcribe_one(save_path)
+                if lang not in list(lang2token.keys()):
+                    print(f"{lang} not supported, ignoring\n")
+                    continue
+                text = "ZH|" + text + "\n"#
+                #text = lang2token[lang] + text + lang2token[lang] + "\n"
+                speaker_annos.append(save_path + "|" + speaker + "|" + text)
+                
+                processed_files += 1
+                print(f"Processed: {processed_files}/{total_files}")
+            except:
+                continue
+
+    # # clean annotation
+    # import argparse
+    # import text
+    # from utils import load_filepaths_and_text
+    # for i, line in enumerate(speaker_annos):
+    #     path, sid, txt = line.split("|")
+    #     cleaned_text = text._clean_text(txt, ["cjke_cleaners2"])
+    #     cleaned_text += "\n" if not cleaned_text.endswith("\n") else ""
+    #     speaker_annos[i] = path + "|" + sid + "|" + cleaned_text
+    # write into annotation
+    if len(speaker_annos) == 0:
+        print("Warning: no short audios found, this IS expected if you have only uploaded long audios, videos or video links.")
+        print("this IS NOT expected if you have uploaded a zip file of short audios. Please check your file structure or make sure your audio language is supported.")
+    with open("./filelists/short_character_anno.list", 'w', encoding='utf-8') as f:
+        for line in speaker_annos:
+            f.write(line)
+
+    # import json
+    # # generate new config
+    # with open("./configs/finetune_speaker.json", 'r', encoding='utf-8') as f:
+    #     hps = json.load(f)
+    # # modify n_speakers
+    # hps['data']["n_speakers"] = 1000 + len(speaker2id)
+    # # add speaker names
+    # for speaker in speaker_names:
+    #     hps['speakers'][speaker] = speaker2id[speaker]
+    # # save modified config
+    # with open("./configs/modified_finetune_speaker.json", 'w', encoding='utf-8') as f:
+    #     json.dump(hps, f, indent=2)
+    # print("finished")

start.bat

@@ -0,0 +1,2 @@
+set PYTHON=venv\python.exe
+start cmd /k "set PYTHON=%PYTHON%"

text/__init__.py

@@ -0,0 +1,28 @@
+from text.symbols import *
+
+
+_symbol_to_id = {s: i for i, s in enumerate(symbols)}
+
+def cleaned_text_to_sequence(cleaned_text, tones, language):
+  '''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
+    Args:
+      text: string to convert to a sequence
+    Returns:
+      List of integers corresponding to the symbols in the text
+  '''
+  phones = [_symbol_to_id[symbol] for symbol in cleaned_text]
+  tone_start = language_tone_start_map[language]
+  tones = [i + tone_start for i in tones]
+  lang_id = language_id_map[language]
+  lang_ids = [lang_id for i in phones]
+  return phones, tones, lang_ids
+
+def get_bert(norm_text, word2ph, language):
+  from .chinese_bert import get_bert_feature as zh_bert
+  from .english_bert_mock import get_bert_feature as en_bert
+  lang_bert_func_map = {
+    'ZH': zh_bert,
+    'EN': en_bert
+  }
+  bert = lang_bert_func_map[language](norm_text, word2ph)
+  return bert

text/chinese.py

@@ -0,0 +1,193 @@
+import os
+import re
+
+import cn2an
+from pypinyin import lazy_pinyin, Style
+
+from text import symbols
+from text.symbols import punctuation
+from text.tone_sandhi import ToneSandhi
+
+current_file_path = os.path.dirname(__file__)
+pinyin_to_symbol_map = {line.split("\t")[0]: line.strip().split("\t")[1] for line in
+                        open(os.path.join(current_file_path, 'opencpop-strict.txt')).readlines()}
+
+import jieba.posseg as psg
+
+
+rep_map = {
+    '：': ',',
+    '；': ',',
+    '，': ',',
+    '。': '.',
+    '！': '!',
+    '？': '?',
+    '\n': '.',
+    "·": ",",
+    '、': ",",
+    '...': '…',
+    '$': '.',
+    '“': "'",
+    '”': "'",
+    '‘': "'",
+    '’': "'",
+    '（': "'",
+    '）': "'",
+    '(': "'",
+    ')': "'",
+    '《': "'",
+    '》': "'",
+    '【': "'",
+    '】': "'",
+    '[': "'",
+    ']': "'",
+    '—': "-",
+    '～': "-",
+    '~': "-",
+    '「': "'",
+    '」': "'",
+
+}
+
+tone_modifier = ToneSandhi()
+
+def replace_punctuation(text):
+    text = text.replace("嗯", "恩").replace("呣","母")
+    pattern = re.compile('|'.join(re.escape(p) for p in rep_map.keys()))
+
+    replaced_text = pattern.sub(lambda x: rep_map[x.group()], text)
+
+    replaced_text = re.sub(r'[^\u4e00-\u9fa5'+"".join(punctuation)+r']+', '', replaced_text)
+
+    return replaced_text
+
+def g2p(text):
+    pattern = r'(?<=[{0}])\s*'.format(''.join(punctuation))
+    sentences = [i for i in re.split(pattern, text) if i.strip()!='']
+    phones, tones, word2ph = _g2p(sentences)
+    assert sum(word2ph) == len(phones)
+    assert len(word2ph) == len(text) #Sometimes it will crash,you can add a try-catch.
+    phones = ['_'] + phones + ["_"]
+    tones = [0] + tones + [0]
+    word2ph = [1] + word2ph + [1]
+    return phones, tones, word2ph
+
+
+def _get_initials_finals(word):
+    initials = []
+    finals = []
+    orig_initials = lazy_pinyin(
+        word, neutral_tone_with_five=True, style=Style.INITIALS)
+    orig_finals = lazy_pinyin(
+        word, neutral_tone_with_five=True, style=Style.FINALS_TONE3)
+    for c, v in zip(orig_initials, orig_finals):
+        initials.append(c)
+        finals.append(v)
+    return initials, finals
+
+
+def _g2p(segments):
+    phones_list = []
+    tones_list = []
+    word2ph = []
+    for seg in segments:
+        pinyins = []
+        # Replace all English words in the sentence
+        seg = re.sub('[a-zA-Z]+', '', seg)
+        seg_cut = psg.lcut(seg)
+        initials = []
+        finals = []
+        seg_cut = tone_modifier.pre_merge_for_modify(seg_cut)
+        for word, pos in seg_cut:
+            if pos == 'eng':
+                continue
+            sub_initials, sub_finals = _get_initials_finals(word)
+            sub_finals = tone_modifier.modified_tone(word, pos,
+                                                          sub_finals)
+            initials.append(sub_initials)
+            finals.append(sub_finals)
+
+            # assert len(sub_initials) == len(sub_finals) == len(word)
+        initials = sum(initials, [])
+        finals = sum(finals, [])
+        #
+        for c, v in zip(initials, finals):
+            raw_pinyin = c+v
+            # NOTE: post process for pypinyin outputs
+            # we discriminate i, ii and iii
+            if c == v:
+                assert c in punctuation
+                phone = [c]
+                tone = '0'
+                word2ph.append(1)
+            else:
+                v_without_tone = v[:-1]
+                tone = v[-1]
+
+                pinyin = c+v_without_tone
+                assert tone in '12345'
+
+                if c:
+                    # 多音节
+                    v_rep_map = {
+                        "uei": 'ui',
+                        'iou': 'iu',
+                        'uen': 'un',
+                    }
+                    if v_without_tone in v_rep_map.keys():
+                        pinyin = c+v_rep_map[v_without_tone]
+                else:
+                    # 单音节
+                    pinyin_rep_map = {
+                        'ing': 'ying',
+                        'i': 'yi',
+                        'in': 'yin',
+                        'u': 'wu',
+                    }
+                    if pinyin in pinyin_rep_map.keys():
+                        pinyin = pinyin_rep_map[pinyin]
+                    else:
+                        single_rep_map = {
+                            'v': 'yu',
+                            'e': 'e',
+                            'i': 'y',
+                            'u': 'w',
+                        }
+                        if pinyin[0] in single_rep_map.keys():
+                            pinyin = single_rep_map[pinyin[0]]+pinyin[1:]
+
+                assert pinyin in pinyin_to_symbol_map.keys(), (pinyin, seg, raw_pinyin)
+                phone = pinyin_to_symbol_map[pinyin].split(' ')
+                word2ph.append(len(phone))
+
+            phones_list += phone
+            tones_list += [int(tone)] * len(phone)
+    return phones_list, tones_list, word2ph
+
+
+
+def text_normalize(text):
+    numbers = re.findall(r'\d+(?:\.?\d+)?', text)
+    for number in numbers:
+        text = text.replace(number, cn2an.an2cn(number), 1)
+    text = replace_punctuation(text)
+    return text
+
+def get_bert_feature(text, word2ph):
+    from  text import chinese_bert
+    return chinese_bert.get_bert_feature(text, word2ph)
+
+if __name__ == '__main__':
+    from text.chinese_bert import get_bert_feature
+    text = "啊！但是《原神》是由,米哈\游自主，  [研发]的一款全.新开放世界.冒险游戏"
+    text = text_normalize(text)
+    print(text)
+    phones, tones, word2ph = g2p(text)
+    bert = get_bert_feature(text, word2ph)
+
+    print(phones, tones, word2ph, bert.shape)
+
+
+# # 示例用法
+# text = "这是一个示例文本：,你好！这是一个测试...."
+# print(g2p_paddle(text))  # 输出: 这是一个示例文本你好这是一个测试

text/chinese_bert.py

@@ -0,0 +1,50 @@
+import torch
+from transformers import AutoTokenizer, AutoModelForMaskedLM
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+tokenizer = AutoTokenizer.from_pretrained("./bert/chinese-roberta-wwm-ext-large")
+model = AutoModelForMaskedLM.from_pretrained("./bert/chinese-roberta-wwm-ext-large").to(device)
+
+def get_bert_feature(text, word2ph):
+    with torch.no_grad():
+        inputs = tokenizer(text, return_tensors='pt')
+        for i in inputs:
+            inputs[i] = inputs[i].to(device)
+        res = model(**inputs, output_hidden_states=True)
+        res = torch.cat(res['hidden_states'][-3:-2], -1)[0].cpu()
+
+    assert len(word2ph) == len(text)+2
+    word2phone = word2ph
+    phone_level_feature = []
+    for i in range(len(word2phone)):
+        repeat_feature = res[i].repeat(word2phone[i], 1)
+        phone_level_feature.append(repeat_feature)
+
+    phone_level_feature = torch.cat(phone_level_feature, dim=0)
+
+
+    return phone_level_feature.T
+
+if __name__ == '__main__':
+    # feature = get_bert_feature('你好,我是说的道理。')
+    import torch
+
+    word_level_feature = torch.rand(38, 1024)  # 12个词,每个词1024维特征
+    word2phone = [1, 2, 1, 2, 2, 1, 2, 2, 1, 2, 2, 1, 2, 2, 2, 2, 2, 1, 1, 2, 2, 1, 2, 2, 2, 2, 1, 2, 2, 2, 2, 2, 1, 2, 2, 2, 2, 1]
+
+    # 计算总帧数
+    total_frames = sum(word2phone)
+    print(word_level_feature.shape)
+    print(word2phone)
+    phone_level_feature = []
+    for i in range(len(word2phone)):
+        print(word_level_feature[i].shape)
+
+        # 对每个词重复word2phone[i]次
+        repeat_feature = word_level_feature[i].repeat(word2phone[i], 1)
+        phone_level_feature.append(repeat_feature)
+
+    phone_level_feature = torch.cat(phone_level_feature, dim=0)
+    print(phone_level_feature.shape)  # torch.Size([36, 1024])
+

text/cleaner.py

@@ -0,0 +1,27 @@
+from text import chinese, cleaned_text_to_sequence
+
+
+language_module_map = {
+    'ZH': chinese
+}
+
+
+def clean_text(text, language):
+    language_module = language_module_map[language]
+    norm_text = language_module.text_normalize(text)
+    phones, tones, word2ph = language_module.g2p(norm_text)
+    return norm_text, phones, tones, word2ph
+
+def clean_text_bert(text, language):
+    language_module = language_module_map[language]
+    norm_text = language_module.text_normalize(text)
+    phones, tones, word2ph = language_module.g2p(norm_text)
+    bert = language_module.get_bert_feature(norm_text, word2ph)
+    return phones, tones, bert
+
+def text_to_sequence(text, language):
+    norm_text, phones, tones, word2ph = clean_text(text, language)
+    return cleaned_text_to_sequence(phones, tones, language)
+
+if __name__ == '__main__':
+    pass

text/cmudict_cache.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9b21b20325471934ba92f2e4a5976989e7d920caa32e7a286eacb027d197949
+size 6212655

text/english.py

@@ -0,0 +1,138 @@
+import pickle
+import os
+import re
+from g2p_en import G2p
+from string import punctuation
+
+from text import symbols
+
+current_file_path = os.path.dirname(__file__)
+CMU_DICT_PATH = os.path.join(current_file_path, 'cmudict.rep')
+CACHE_PATH = os.path.join(current_file_path, 'cmudict_cache.pickle')
+_g2p = G2p()
+
+arpa = {'AH0', 'S', 'AH1', 'EY2', 'AE2', 'EH0', 'OW2', 'UH0', 'NG', 'B', 'G', 'AY0', 'M', 'AA0', 'F', 'AO0', 'ER2', 'UH1', 'IY1', 'AH2', 'DH', 'IY0', 'EY1', 'IH0', 'K', 'N', 'W', 'IY2', 'T', 'AA1', 'ER1', 'EH2', 'OY0', 'UH2', 'UW1', 'Z', 'AW2', 'AW1', 'V', 'UW2', 'AA2', 'ER', 'AW0', 'UW0', 'R', 'OW1', 'EH1', 'ZH', 'AE0', 'IH2', 'IH', 'Y', 'JH', 'P', 'AY1', 'EY0', 'OY2', 'TH', 'HH', 'D', 'ER0', 'CH', 'AO1', 'AE1', 'AO2', 'OY1', 'AY2', 'IH1', 'OW0', 'L', 'SH'}
+
+
+def post_replace_ph(ph):
+    rep_map = {
+        '：': ',',
+        '；': ',',
+        '，': ',',
+        '。': '.',
+        '！': '!',
+        '？': '?',
+        '\n': '.',
+        "·": ",",
+        '、': ",",
+        '...': '…',
+        'v': "V"
+    }
+    if ph in rep_map.keys():
+        ph = rep_map[ph]
+    if ph in symbols:
+        return ph
+    if ph not in symbols:
+        ph = 'UNK'
+    return ph
+
+def read_dict():
+    g2p_dict = {}
+    start_line = 49
+    with open(CMU_DICT_PATH) as f:
+        line = f.readline()
+        line_index = 1
+        while line:
+            if line_index >= start_line:
+                line = line.strip()
+                word_split = line.split('  ')
+                word = word_split[0]
+
+                syllable_split = word_split[1].split(' - ')
+                g2p_dict[word] = []
+                for syllable in syllable_split:
+                    phone_split = syllable.split(' ')
+                    g2p_dict[word].append(phone_split)
+
+            line_index = line_index + 1
+            line = f.readline()
+
+    return g2p_dict
+
+
+def cache_dict(g2p_dict, file_path):
+    with open(file_path, 'wb') as pickle_file:
+        pickle.dump(g2p_dict, pickle_file)
+
+
+def get_dict():
+    if os.path.exists(CACHE_PATH):
+        with open(CACHE_PATH, 'rb') as pickle_file:
+            g2p_dict = pickle.load(pickle_file)
+    else:
+        g2p_dict = read_dict()
+        cache_dict(g2p_dict, CACHE_PATH)
+
+    return g2p_dict
+
+eng_dict = get_dict()
+
+def refine_ph(phn):
+    tone = 0
+    if re.search(r'\d$', phn):
+        tone = int(phn[-1]) + 1
+        phn = phn[:-1]
+    return phn.lower(), tone
+
+def refine_syllables(syllables):
+    tones = []
+    phonemes = []
+    for phn_list in syllables:
+        for i in range(len(phn_list)):
+            phn = phn_list[i]
+            phn, tone = refine_ph(phn)
+            phonemes.append(phn)
+            tones.append(tone)
+    return phonemes, tones
+
+
+def text_normalize(text):
+    # todo: eng text normalize
+    return text
+
+def g2p(text):
+
+    phones = []
+    tones = []
+    words = re.split(r"([,;.\-\?\!\s+])", text)
+    for w in words:
+        if w.upper() in eng_dict:
+            phns, tns = refine_syllables(eng_dict[w.upper()])
+            phones += phns
+            tones += tns
+        else:
+            phone_list = list(filter(lambda p: p != " ", _g2p(w)))
+            for ph in phone_list:
+                if ph in arpa:
+                    ph, tn = refine_ph(ph)
+                    phones.append(ph)
+                    tones.append(tn)
+                else:
+                    phones.append(ph)
+                    tones.append(0)
+    # todo: implement word2ph
+    word2ph = [1 for i in phones]
+
+    phones = [post_replace_ph(i) for i in phones]
+    return phones, tones, word2ph
+
+if __name__ == "__main__":
+    # print(get_dict())
+    # print(eng_word_to_phoneme("hello"))
+    print(g2p("In this paper, we propose 1 DSPGAN, a GAN-based universal vocoder."))
+    # all_phones = set()
+    # for k, syllables in eng_dict.items():
+    #     for group in syllables:
+    #         for ph in group:
+    #             all_phones.add(ph)
+    # print(all_phones)

text/english_bert_mock.py

@@ -0,0 +1,5 @@
+import torch
+
+
+def get_bert_feature(norm_text, word2ph):
+    return torch.zeros(1024, sum(word2ph))

text/japanese.py

@@ -0,0 +1,104 @@
+# modified from https://github.com/CjangCjengh/vits/blob/main/text/japanese.py
+import re
+import sys
+
+import pyopenjtalk
+
+from text import symbols
+
+# Regular expression matching Japanese without punctuation marks:
+_japanese_characters = re.compile(
+    r'[A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]')
+
+# Regular expression matching non-Japanese characters or punctuation marks:
+_japanese_marks = re.compile(
+    r'[^A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]')
+
+# List of (symbol, Japanese) pairs for marks:
+_symbols_to_japanese = [(re.compile('%s' % x[0]), x[1]) for x in [
+    ('％', 'パーセント')
+]]
+
+
+# List of (consonant, sokuon) pairs:
+_real_sokuon = [(re.compile('%s' % x[0]), x[1]) for x in [
+    (r'Q([↑↓]*[kg])', r'k#\1'),
+    (r'Q([↑↓]*[tdjʧ])', r't#\1'),
+    (r'Q([↑↓]*[sʃ])', r's\1'),
+    (r'Q([↑↓]*[pb])', r'p#\1')
+]]
+
+# List of (consonant, hatsuon) pairs:
+_real_hatsuon = [(re.compile('%s' % x[0]), x[1]) for x in [
+    (r'N([↑↓]*[pbm])', r'm\1'),
+    (r'N([↑↓]*[ʧʥj])', r'n^\1'),
+    (r'N([↑↓]*[tdn])', r'n\1'),
+    (r'N([↑↓]*[kg])', r'ŋ\1')
+]]
+
+
+
+def post_replace_ph(ph):
+    rep_map = {
+        '：': ',',
+        '；': ',',
+        '，': ',',
+        '。': '.',
+        '！': '!',
+        '？': '?',
+        '\n': '.',
+        "·": ",",
+        '、': ",",
+        '...': '…',
+        'v': "V"
+    }
+    if ph in rep_map.keys():
+        ph = rep_map[ph]
+    if ph in symbols:
+        return ph
+    if ph not in symbols:
+        ph = 'UNK'
+    return ph
+
+def symbols_to_japanese(text):
+    for regex, replacement in _symbols_to_japanese:
+        text = re.sub(regex, replacement, text)
+    return text
+
+
+def preprocess_jap(text):
+    '''Reference https://r9y9.github.io/ttslearn/latest/notebooks/ch10_Recipe-Tacotron.html'''
+    text = symbols_to_japanese(text)
+    sentences = re.split(_japanese_marks, text)
+    marks = re.findall(_japanese_marks, text)
+    text = []
+    for i, sentence in enumerate(sentences):
+        if re.match(_japanese_characters, sentence):
+            p = pyopenjtalk.g2p(sentence)
+            text += p.split(" ")
+
+        if i < len(marks):
+            text += [marks[i].replace(' ', '')]
+    return text
+
+def text_normalize(text):
+    # todo: jap text normalize
+    return text
+
+def g2p(norm_text):
+    phones = preprocess_jap(norm_text)
+    phones = [post_replace_ph(i) for i in phones]
+    # todo: implement tones and word2ph
+    tones = [0 for i in phones]
+    word2ph = [1 for i in phones]
+    return phones, tones, word2ph
+
+
+if __name__ == '__main__':
+    for line in open("../../../Downloads/transcript_utf8.txt").readlines():
+        text = line.split(":")[1]
+        phones, tones, word2ph = g2p(text)
+        for p in phones:
+            if p == "z":
+                print(text, phones)
+                sys.exit(0)

text/opencpop-strict.txt

@@ -0,0 +1,429 @@
+a	AA a
+ai	AA ai
+an	AA an
+ang	AA ang
+ao	AA ao
+ba	b a
+bai	b ai
+ban	b an
+bang	b ang
+bao	b ao
+bei	b ei
+ben	b en
+beng	b eng
+bi	b i
+bian	b ian
+biao	b iao
+bie	b ie
+bin	b in
+bing	b ing
+bo	b o
+bu	b u
+ca	c a
+cai	c ai
+can	c an
+cang	c ang
+cao	c ao
+ce	c e
+cei	c ei
+cen	c en
+ceng	c eng
+cha	ch a
+chai	ch ai
+chan	ch an
+chang	ch ang
+chao	ch ao
+che	ch e
+chen	ch en
+cheng	ch eng
+chi	ch ir
+chong	ch ong
+chou	ch ou
+chu	ch u
+chua	ch ua
+chuai	ch uai
+chuan	ch uan
+chuang	ch uang
+chui	ch ui
+chun	ch un
+chuo	ch uo
+ci	c i0
+cong	c ong
+cou	c ou
+cu	c u
+cuan	c uan
+cui	c ui
+cun	c un
+cuo	c uo
+da	d a
+dai	d ai
+dan	d an
+dang	d ang
+dao	d ao
+de	d e
+dei	d ei
+den	d en
+deng	d eng
+di	d i
+dia	d ia
+dian	d ian
+diao	d iao
+die	d ie
+ding	d ing
+diu	d iu
+dong	d ong
+dou	d ou
+du	d u
+duan	d uan
+dui	d ui
+dun	d un
+duo	d uo
+e	EE e
+ei	EE ei
+en	EE en
+eng	EE eng
+er	EE er
+fa	f a
+fan	f an
+fang	f ang
+fei	f ei
+fen	f en
+feng	f eng
+fo	f o
+fou	f ou
+fu	f u
+ga	g a
+gai	g ai
+gan	g an
+gang	g ang
+gao	g ao
+ge	g e
+gei	g ei
+gen	g en
+geng	g eng
+gong	g ong
+gou	g ou
+gu	g u
+gua	g ua
+guai	g uai
+guan	g uan
+guang	g uang
+gui	g ui
+gun	g un
+guo	g uo
+ha	h a
+hai	h ai
+han	h an
+hang	h ang
+hao	h ao
+he	h e
+hei	h ei
+hen	h en
+heng	h eng
+hong	h ong
+hou	h ou
+hu	h u
+hua	h ua
+huai	h uai
+huan	h uan
+huang	h uang
+hui	h ui
+hun	h un
+huo	h uo
+ji	j i
+jia	j ia
+jian	j ian
+jiang	j iang
+jiao	j iao
+jie	j ie
+jin	j in
+jing	j ing
+jiong	j iong
+jiu	j iu
+ju	j v
+jv	j v
+juan	j van
+jvan	j van
+jue	j ve
+jve	j ve
+jun	j vn
+jvn	j vn
+ka	k a
+kai	k ai
+kan	k an
+kang	k ang
+kao	k ao
+ke	k e
+kei	k ei
+ken	k en
+keng	k eng
+kong	k ong
+kou	k ou
+ku	k u
+kua	k ua
+kuai	k uai
+kuan	k uan
+kuang	k uang
+kui	k ui
+kun	k un
+kuo	k uo
+la	l a
+lai	l ai
+lan	l an
+lang	l ang
+lao	l ao
+le	l e
+lei	l ei
+leng	l eng
+li	l i
+lia	l ia
+lian	l ian
+liang	l iang
+liao	l iao
+lie	l ie
+lin	l in
+ling	l ing
+liu	l iu
+lo	l o
+long	l ong
+lou	l ou
+lu	l u
+luan	l uan
+lun	l un
+luo	l uo
+lv	l v
+lve	l ve
+ma	m a
+mai	m ai
+man	m an
+mang	m ang
+mao	m ao
+me	m e
+mei	m ei
+men	m en
+meng	m eng
+mi	m i
+mian	m ian
+miao	m iao
+mie	m ie
+min	m in
+ming	m ing
+miu	m iu
+mo	m o
+mou	m ou
+mu	m u
+na	n a
+nai	n ai
+nan	n an
+nang	n ang
+nao	n ao
+ne	n e
+nei	n ei
+nen	n en
+neng	n eng
+ni	n i
+nian	n ian
+niang	n iang
+niao	n iao
+nie	n ie
+nin	n in
+ning	n ing
+niu	n iu
+nong	n ong
+nou	n ou
+nu	n u
+nuan	n uan
+nun	n un
+nuo	n uo
+nv	n v
+nve	n ve
+o	OO o
+ou	OO ou
+pa	p a
+pai	p ai
+pan	p an
+pang	p ang
+pao	p ao
+pei	p ei
+pen	p en
+peng	p eng
+pi	p i
+pian	p ian
+piao	p iao
+pie	p ie
+pin	p in
+ping	p ing
+po	p o
+pou	p ou
+pu	p u
+qi	q i
+qia	q ia
+qian	q ian
+qiang	q iang
+qiao	q iao
+qie	q ie
+qin	q in
+qing	q ing
+qiong	q iong
+qiu	q iu
+qu	q v
+qv	q v
+quan	q van
+qvan	q van
+que	q ve
+qve	q ve
+qun	q vn
+qvn	q vn
+ran	r an
+rang	r ang
+rao	r ao
+re	r e
+ren	r en
+reng	r eng
+ri	r ir
+rong	r ong
+rou	r ou
+ru	r u
+rua	r ua
+ruan	r uan
+rui	r ui
+run	r un
+ruo	r uo
+sa	s a
+sai	s ai
+san	s an
+sang	s ang
+sao	s ao
+se	s e
+sen	s en
+seng	s eng
+sha	sh a
+shai	sh ai
+shan	sh an
+shang	sh ang
+shao	sh ao
+she	sh e
+shei	sh ei
+shen	sh en
+sheng	sh eng
+shi	sh ir
+shou	sh ou
+shu	sh u
+shua	sh ua
+shuai	sh uai
+shuan	sh uan
+shuang	sh uang
+shui	sh ui
+shun	sh un
+shuo	sh uo
+si	s i0
+song	s ong
+sou	s ou
+su	s u
+suan	s uan
+sui	s ui
+sun	s un
+suo	s uo
+ta	t a
+tai	t ai
+tan	t an
+tang	t ang
+tao	t ao
+te	t e
+tei	t ei
+teng	t eng
+ti	t i
+tian	t ian
+tiao	t iao
+tie	t ie
+ting	t ing
+tong	t ong
+tou	t ou
+tu	t u
+tuan	t uan
+tui	t ui
+tun	t un
+tuo	t uo
+wa	w a
+wai	w ai
+wan	w an
+wang	w ang
+wei	w ei
+wen	w en
+weng	w eng
+wo	w o
+wu	w u
+xi	x i
+xia	x ia
+xian	x ian
+xiang	x iang
+xiao	x iao
+xie	x ie
+xin	x in
+xing	x ing
+xiong	x iong
+xiu	x iu
+xu	x v
+xv	x v
+xuan	x van
+xvan	x van
+xue	x ve
+xve	x ve
+xun	x vn
+xvn	x vn
+ya	y a
+yan	y En
+yang	y ang
+yao	y ao
+ye	y E
+yi	y i
+yin	y in
+ying	y ing
+yo	y o
+yong	y ong
+you	y ou
+yu	y v
+yv	y v
+yuan	y van
+yvan	y van
+yue	y ve
+yve	y ve
+yun	y vn
+yvn	y vn
+za	z a
+zai	z ai
+zan	z an
+zang	z ang
+zao	z ao
+ze	z e
+zei	z ei
+zen	z en
+zeng	z eng
+zha	zh a
+zhai	zh ai
+zhan	zh an
+zhang	zh ang
+zhao	zh ao
+zhe	zh e
+zhei	zh ei
+zhen	zh en
+zheng	zh eng
+zhi	zh ir
+zhong	zh ong
+zhou	zh ou
+zhu	zh u
+zhua	zh ua
+zhuai	zh uai
+zhuan	zh uan
+zhuang	zh uang
+zhui	zh ui
+zhun	zh un
+zhuo	zh uo
+zi	z i0
+zong	z ong
+zou	z ou
+zu	z u
+zuan	z uan
+zui	z ui
+zun	z un
+zuo	z uo

text/symbols.py

@@ -0,0 +1,51 @@
+punctuation = ['!', '?', '…', ",", ".", "'", '-']
+pu_symbols = punctuation + ["SP", "UNK"]
+pad = '_'
+
+# chinese
+zh_symbols = ['E', 'En', 'a', 'ai', 'an', 'ang', 'ao', 'b', 'c', 'ch', 'd', 'e', 'ei', 'en', 'eng', 'er', 'f', 'g', 'h',
+       'i', 'i0', 'ia', 'ian', 'iang', 'iao', 'ie', 'in', 'ing', 'iong', 'ir', 'iu', 'j', 'k', 'l', 'm', 'n', 'o',
+       'ong',
+       'ou', 'p', 'q', 'r', 's', 'sh', 't', 'u', 'ua', 'uai', 'uan', 'uang', 'ui', 'un', 'uo', 'v', 'van', 've', 'vn',
+       'w', 'x', 'y', 'z', 'zh',
+        "AA", "EE", "OO"]
+num_zh_tones = 6
+
+# japanese
+ja_symbols = ['I', 'N', 'U', 'a', 'b', 'by', 'ch', 'cl', 'd', 'dy', 'e', 'f', 'g', 'gy', 'h', 'hy', 'i', 'j', 'k', 'ky',
+              'm', 'my', 'n', 'ny', 'o', 'p', 'py', 'r', 'ry', 's', 'sh', 't', 'ts', 'u', 'V', 'w', 'y', 'z']
+num_ja_tones = 1
+
+# English
+en_symbols = ['aa', 'ae', 'ah', 'ao', 'aw', 'ay', 'b', 'ch', 'd', 'dh', 'eh', 'er', 'ey', 'f', 'g', 'hh', 'ih', 'iy',
+              'jh', 'k', 'l', 'm', 'n', 'ng', 'ow', 'oy', 'p', 'r', 's',
+              'sh', 't', 'th', 'uh', 'uw', 'V', 'w', 'y', 'z', 'zh']
+num_en_tones = 4
+
+# combine all symbols
+normal_symbols = sorted(set(zh_symbols + ja_symbols + en_symbols))
+symbols = [pad] + normal_symbols + pu_symbols
+sil_phonemes_ids = [symbols.index(i) for i in pu_symbols]
+
+# combine all tones
+num_tones = num_zh_tones + num_ja_tones + num_en_tones
+
+# language maps
+language_id_map = {
+    'ZH': 0,
+    "JA": 1,
+    "EN": 2
+}
+num_languages = len(language_id_map.keys())
+
+language_tone_start_map = {
+    'ZH': 0,
+    "JA": num_zh_tones,
+    "EN": num_zh_tones + num_ja_tones
+}
+
+if __name__ == '__main__':
+    a = set(zh_symbols)
+    b = set(en_symbols)
+    print(sorted(a&b))
+

text/tone_sandhi.py

@@ -0,0 +1,351 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import List
+from typing import Tuple
+
+import jieba
+from pypinyin import lazy_pinyin
+from pypinyin import Style
+
+
+class ToneSandhi():
+    def __init__(self):
+        self.must_neural_tone_words = {
+            '麻烦', '麻利', '鸳鸯', '高粱', '骨头', '骆驼', '马虎', '首饰', '馒头', '馄饨', '风筝',
+            '难为', '队伍', '阔气', '闺女', '门道', '锄头', '铺盖', '铃铛', '铁匠', '钥匙', '里脊',
+            '里头', '部分', '那么', '道士', '造化', '迷糊', '连累', '这么', '这个', '运气', '过去',
+            '软和', '转悠', '踏实', '跳蚤', '跟头', '趔趄', '财主', '豆腐', '讲究', '记性', '记号',
+            '认识', '规矩', '见识', '裁缝', '补丁', '衣裳', '衣服', '衙门', '街坊', '行李', '行当',
+            '蛤蟆', '蘑菇', '薄荷', '葫芦', '葡萄', '萝卜', '荸荠', '苗条', '苗头', '苍蝇', '芝麻',
+            '舒服', '舒坦', '舌头', '自在', '膏药', '脾气', '脑袋', '脊梁', '能耐', '胳膊', '胭脂',
+            '胡萝', '胡琴', '胡同', '聪明', '耽误', '耽搁', '耷拉', '耳朵', '老爷', '老实', '老婆',
+            '老头', '老太', '翻腾', '罗嗦', '罐头', '编辑', '结实', '红火', '累赘', '糨糊', '糊涂',
+            '精神', '粮食', '簸箕', '篱笆', '算计', '算盘', '答应', '笤帚', '笑语', '笑话', '窟窿',
+            '窝囊', '窗户', '稳当', '稀罕', '称呼', '秧歌', '秀气', '秀才', '福气', '祖宗', '砚台',
+            '码头', '石榴', '石头', '石匠', '知识', '眼睛', '眯缝', '眨巴', '眉毛', '相声', '盘算',
+            '白净', '痢疾', '痛快', '疟疾', '疙瘩', '疏忽', '畜生', '生意', '甘蔗', '琵琶', '琢磨',
+            '琉璃', '玻璃', '玫瑰', '玄乎', '狐狸', '状元', '特务', '牲口', '牙碜', '牌楼', '爽快',
+            '爱人', '热闹', '烧饼', '烟筒', '烂糊', '点心', '炊帚', '灯笼', '火候', '漂亮', '滑溜',
+            '溜达', '温和', '清楚', '消息', '浪头', '活泼', '比方', '正经', '欺负', '模糊', '槟榔',
+            '棺材', '棒槌', '棉花', '核桃', '栅栏', '柴火', '架势', '枕头', '枇杷', '机灵', '本事',
+            '木头', '木匠', '朋友', '月饼', '月亮', '暖和', '明白', '时候', '新鲜', '故事', '收拾',
+            '收成', '提防', '挖苦', '挑剔', '指甲', '指头', '拾掇', '拳头', '拨弄', '招牌', '招呼',
+            '抬举', '护士', '折腾', '扫帚', '打量', '打算', '打点', '打扮', '打听', '打发', '扎实',
+            '扁担', '戒指', '懒得', '意识', '意思', '情形', '悟性', '怪物', '思量', '怎么', '念头',
+            '念叨', '快活', '忙活', '志气', '心思', '得罪', '张罗', '弟兄', '开通', '应酬', '庄稼',
+            '干事', '帮手', '帐篷', '希罕', '师父', '师傅', '巴结', '巴掌', '差事', '工夫', '岁数',
+            '屁股', '尾巴', '少爷', '小气', '小伙', '将就', '对头', '对付', '寡妇', '家伙', '客气',
+            '实在', '官司', '学问', '学生', '字号', '嫁妆', '媳妇', '媒人', '婆家', '娘家', '委屈',
+            '姑娘', '姐夫', '妯娌', '妥当', '妖精', '奴才', '女婿', '头发', '太阳', '大爷', '大方',
+            '大意', '大夫', '多少', '多么', '外甥', '壮实', '地道', '地方', '在乎', '困难', '嘴巴',
+            '嘱咐', '嘟囔', '嘀咕', '喜欢', '喇嘛', '喇叭', '商量', '唾沫', '哑巴', '哈欠', '哆嗦',
+            '咳嗽', '和尚', '告诉', '告示', '含糊', '吓唬', '后头', '名字', '名堂', '合同', '吆喝',
+            '叫唤', '口袋', '厚道', '厉害', '千斤', '包袱', '包涵', '匀称', '勤快', '动静', '动弹',
+            '功夫', '力气', '前头', '刺猬', '刺激', '别扭', '利落', '利索', '利害', '分析', '出息',
+            '凑合', '凉快', '冷战', '冤枉', '冒失', '养活', '关系', '先生', '兄弟', '便宜', '使唤',
+            '佩服', '作坊', '体面', '位置', '似的', '伙计', '休息', '什么', '人家', '亲戚', '亲家',
+            '交情', '云彩', '事情', '买卖', '主意', '丫头', '丧气', '两口', '东西', '东家', '世故',
+            '不由', '不在', '下水', '下巴', '上头', '上司', '丈夫', '丈人', '一辈', '那个', '菩萨',
+            '父亲', '母亲', '咕噜', '邋遢', '费用', '冤家', '甜头', '介绍', '荒唐', '大人', '泥鳅',
+            '幸福', '熟悉', '计划', '扑腾', '蜡烛', '姥爷', '照顾', '喉咙', '吉他', '弄堂', '蚂蚱',
+            '凤凰', '拖沓', '寒碜', '糟蹋', '倒腾', '报复', '逻辑', '盘缠', '喽啰', '牢骚', '咖喱',
+            '扫把', '惦记'
+        }
+        self.must_not_neural_tone_words = {
+            "男子", "女子", "分子", "原子", "量子", "莲子", "石子", "瓜子", "电子", "人人", "虎虎"
+        }
+        self.punc = "：，；。？！“”‘’':,;.?!"
+
+    # the meaning of jieba pos tag: https://blog.csdn.net/weixin_44174352/article/details/113731041
+    # e.g.
+    # word: "家里"
+    # pos: "s"
+    # finals: ['ia1', 'i3']
+    def _neural_sandhi(self, word: str, pos: str,
+                       finals: List[str]) -> List[str]:
+
+        # reduplication words for n. and v. e.g. 奶奶, 试试, 旺旺
+        for j, item in enumerate(word):
+            if j - 1 >= 0 and item == word[j - 1] and pos[0] in {
+                    "n", "v", "a"
+            } and word not in self.must_not_neural_tone_words:
+                finals[j] = finals[j][:-1] + "5"
+        ge_idx = word.find("个")
+        if len(word) >= 1 and word[-1] in "吧呢啊呐噻嘛吖嗨呐哦哒额滴哩哟喽啰耶喔诶":
+            finals[-1] = finals[-1][:-1] + "5"
+        elif len(word) >= 1 and word[-1] in "的地得":
+            finals[-1] = finals[-1][:-1] + "5"
+        # e.g. 走了, 看着, 去过
+        # elif len(word) == 1 and word in "了着过" and pos in {"ul", "uz", "ug"}:
+        #     finals[-1] = finals[-1][:-1] + "5"
+        elif len(word) > 1 and word[-1] in "们子" and pos in {
+                "r", "n"
+        } and word not in self.must_not_neural_tone_words:
+            finals[-1] = finals[-1][:-1] + "5"
+        # e.g. 桌上, 地下, 家里
+        elif len(word) > 1 and word[-1] in "上下里" and pos in {"s", "l", "f"}:
+            finals[-1] = finals[-1][:-1] + "5"
+        # e.g. 上来, 下去
+        elif len(word) > 1 and word[-1] in "来去" and word[-2] in "上下进出回过起开":
+            finals[-1] = finals[-1][:-1] + "5"
+        # 个做量词
+        elif (ge_idx >= 1 and
+              (word[ge_idx - 1].isnumeric() or
+               word[ge_idx - 1] in "几有两半多各整每做是")) or word == '个':
+            finals[ge_idx] = finals[ge_idx][:-1] + "5"
+        else:
+            if word in self.must_neural_tone_words or word[
+                    -2:] in self.must_neural_tone_words:
+                finals[-1] = finals[-1][:-1] + "5"
+
+        word_list = self._split_word(word)
+        finals_list = [finals[:len(word_list[0])], finals[len(word_list[0]):]]
+        for i, word in enumerate(word_list):
+            # conventional neural in Chinese
+            if word in self.must_neural_tone_words or word[
+                    -2:] in self.must_neural_tone_words:
+                finals_list[i][-1] = finals_list[i][-1][:-1] + "5"
+        finals = sum(finals_list, [])
+        return finals
+
+    def _bu_sandhi(self, word: str, finals: List[str]) -> List[str]:
+        # e.g. 看不懂
+        if len(word) == 3 and word[1] == "不":
+            finals[1] = finals[1][:-1] + "5"
+        else:
+            for i, char in enumerate(word):
+                # "不" before tone4 should be bu2, e.g. 不怕
+                if char == "不" and i + 1 < len(word) and finals[i +
+                                                                1][-1] == "4":
+                    finals[i] = finals[i][:-1] + "2"
+        return finals
+
+    def _yi_sandhi(self, word: str, finals: List[str]) -> List[str]:
+        # "一" in number sequences, e.g. 一零零, 二一零
+        if word.find("一") != -1 and all(
+            [item.isnumeric() for item in word if item != "一"]):
+            return finals
+        # "一" between reduplication words shold be yi5, e.g. 看一看
+        elif len(word) == 3 and word[1] == "一" and word[0] == word[-1]:
+            finals[1] = finals[1][:-1] + "5"
+        # when "一" is ordinal word, it should be yi1
+        elif word.startswith("第一"):
+            finals[1] = finals[1][:-1] + "1"
+        else:
+            for i, char in enumerate(word):
+                if char == "一" and i + 1 < len(word):
+                    # "一" before tone4 should be yi2, e.g. 一段
+                    if finals[i + 1][-1] == "4":
+                        finals[i] = finals[i][:-1] + "2"
+                    # "一" before non-tone4 should be yi4, e.g. 一天
+                    else:
+                        # "一" 后面如果是标点，还读一声
+                        if word[i + 1] not in self.punc:
+                            finals[i] = finals[i][:-1] + "4"
+        return finals
+
+    def _split_word(self, word: str) -> List[str]:
+        word_list = jieba.cut_for_search(word)
+        word_list = sorted(word_list, key=lambda i: len(i), reverse=False)
+        first_subword = word_list[0]
+        first_begin_idx = word.find(first_subword)
+        if first_begin_idx == 0:
+            second_subword = word[len(first_subword):]
+            new_word_list = [first_subword, second_subword]
+        else:
+            second_subword = word[:-len(first_subword)]
+            new_word_list = [second_subword, first_subword]
+        return new_word_list
+
+    def _three_sandhi(self, word: str, finals: List[str]) -> List[str]:
+        if len(word) == 2 and self._all_tone_three(finals):
+            finals[0] = finals[0][:-1] + "2"
+        elif len(word) == 3:
+            word_list = self._split_word(word)
+            if self._all_tone_three(finals):
+                #  disyllabic + monosyllabic, e.g. 蒙古/包
+                if len(word_list[0]) == 2:
+                    finals[0] = finals[0][:-1] + "2"
+                    finals[1] = finals[1][:-1] + "2"
+                #  monosyllabic + disyllabic, e.g. 纸/老虎
+                elif len(word_list[0]) == 1:
+                    finals[1] = finals[1][:-1] + "2"
+            else:
+                finals_list = [
+                    finals[:len(word_list[0])], finals[len(word_list[0]):]
+                ]
+                if len(finals_list) == 2:
+                    for i, sub in enumerate(finals_list):
+                        # e.g. 所有/人
+                        if self._all_tone_three(sub) and len(sub) == 2:
+                            finals_list[i][0] = finals_list[i][0][:-1] + "2"
+                        # e.g. 好/喜欢
+                        elif i == 1 and not self._all_tone_three(sub) and finals_list[i][0][-1] == "3" and \
+                                finals_list[0][-1][-1] == "3":
+
+                            finals_list[0][-1] = finals_list[0][-1][:-1] + "2"
+                        finals = sum(finals_list, [])
+        # split idiom into two words who's length is 2
+        elif len(word) == 4:
+            finals_list = [finals[:2], finals[2:]]
+            finals = []
+            for sub in finals_list:
+                if self._all_tone_three(sub):
+                    sub[0] = sub[0][:-1] + "2"
+                finals += sub
+
+        return finals
+
+    def _all_tone_three(self, finals: List[str]) -> bool:
+        return all(x[-1] == "3" for x in finals)
+
+    # merge "不" and the word behind it
+    # if don't merge, "不" sometimes appears alone according to jieba, which may occur sandhi error
+    def _merge_bu(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        last_word = ""
+        for word, pos in seg:
+            if last_word == "不":
+                word = last_word + word
+            if word != "不":
+                new_seg.append((word, pos))
+            last_word = word[:]
+        if last_word == "不":
+            new_seg.append((last_word, 'd'))
+            last_word = ""
+        return new_seg
+
+    # function 1: merge "一" and reduplication words in it's left and right, e.g. "听","一","听" ->"听一听"
+    # function 2: merge single  "一" and the word behind it
+    # if don't merge, "一" sometimes appears alone according to jieba, which may occur sandhi error
+    # e.g.
+    # input seg: [('听', 'v'), ('一', 'm'), ('听', 'v')]
+    # output seg: [['听一听', 'v']]
+    def _merge_yi(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        # function 1
+        for i, (word, pos) in enumerate(seg):
+            if i - 1 >= 0 and word == "一" and i + 1 < len(seg) and seg[i - 1][
+                    0] == seg[i + 1][0] and seg[i - 1][1] == "v":
+                new_seg[i - 1][0] = new_seg[i - 1][0] + "一" + new_seg[i - 1][0]
+            else:
+                if i - 2 >= 0 and seg[i - 1][0] == "一" and seg[i - 2][
+                        0] == word and pos == "v":
+                    continue
+                else:
+                    new_seg.append([word, pos])
+        seg = new_seg
+        new_seg = []
+        # function 2
+        for i, (word, pos) in enumerate(seg):
+            if new_seg and new_seg[-1][0] == "一":
+                new_seg[-1][0] = new_seg[-1][0] + word
+            else:
+                new_seg.append([word, pos])
+        return new_seg
+
+    # the first and the second words are all_tone_three
+    def _merge_continuous_three_tones(
+            self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        sub_finals_list = [
+            lazy_pinyin(
+                word, neutral_tone_with_five=True, style=Style.FINALS_TONE3)
+            for (word, pos) in seg
+        ]
+        assert len(sub_finals_list) == len(seg)
+        merge_last = [False] * len(seg)
+        for i, (word, pos) in enumerate(seg):
+            if i - 1 >= 0 and self._all_tone_three(
+                    sub_finals_list[i - 1]) and self._all_tone_three(
+                        sub_finals_list[i]) and not merge_last[i - 1]:
+                # if the last word is reduplication, not merge, because reduplication need to be _neural_sandhi
+                if not self._is_reduplication(seg[i - 1][0]) and len(
+                        seg[i - 1][0]) + len(seg[i][0]) <= 3:
+                    new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
+                    merge_last[i] = True
+                else:
+                    new_seg.append([word, pos])
+            else:
+                new_seg.append([word, pos])
+
+        return new_seg
+
+    def _is_reduplication(self, word: str) -> bool:
+        return len(word) == 2 and word[0] == word[1]
+
+    # the last char of first word and the first char of second word is tone_three
+    def _merge_continuous_three_tones_2(
+            self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        sub_finals_list = [
+            lazy_pinyin(
+                word, neutral_tone_with_five=True, style=Style.FINALS_TONE3)
+            for (word, pos) in seg
+        ]
+        assert len(sub_finals_list) == len(seg)
+        merge_last = [False] * len(seg)
+        for i, (word, pos) in enumerate(seg):
+            if i - 1 >= 0 and sub_finals_list[i - 1][-1][-1] == "3" and sub_finals_list[i][0][-1] == "3" and not \
+                    merge_last[i - 1]:
+                # if the last word is reduplication, not merge, because reduplication need to be _neural_sandhi
+                if not self._is_reduplication(seg[i - 1][0]) and len(
+                        seg[i - 1][0]) + len(seg[i][0]) <= 3:
+                    new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
+                    merge_last[i] = True
+                else:
+                    new_seg.append([word, pos])
+            else:
+                new_seg.append([word, pos])
+        return new_seg
+
+    def _merge_er(self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        for i, (word, pos) in enumerate(seg):
+            if i - 1 >= 0 and word == "儿" and seg[i-1][0] != "#":
+                new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
+            else:
+                new_seg.append([word, pos])
+        return new_seg
+
+    def _merge_reduplication(
+            self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        new_seg = []
+        for i, (word, pos) in enumerate(seg):
+            if new_seg and word == new_seg[-1][0]:
+                new_seg[-1][0] = new_seg[-1][0] + seg[i][0]
+            else:
+                new_seg.append([word, pos])
+        return new_seg
+
+    def pre_merge_for_modify(
+            self, seg: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
+        seg = self._merge_bu(seg)
+        try:
+            seg = self._merge_yi(seg)
+        except:
+            print("_merge_yi failed")
+        seg = self._merge_reduplication(seg)
+        seg = self._merge_continuous_three_tones(seg)
+        seg = self._merge_continuous_three_tones_2(seg)
+        seg = self._merge_er(seg)
+        return seg
+
+    def modified_tone(self, word: str, pos: str,
+                      finals: List[str]) -> List[str]:
+        finals = self._bu_sandhi(word, finals)
+        finals = self._yi_sandhi(word, finals)
+        finals = self._neural_sandhi(word, pos, finals)
+        finals = self._three_sandhi(word, finals)
+        return finals

train_ms.py

@@ -0,0 +1,402 @@
+import os
+import json
+import argparse
+import itertools
+import math
+import torch
+import shutil
+from torch import nn, optim
+from torch.nn import functional as F
+from torch.utils.data import DataLoader
+from torch.utils.tensorboard import SummaryWriter
+import torch.multiprocessing as mp
+import torch.distributed as dist
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.cuda.amp import autocast, GradScaler
+from tqdm import tqdm
+import logging
+logging.getLogger('numba').setLevel(logging.WARNING)
+import commons
+import utils
+from data_utils import (
+    TextAudioSpeakerLoader,
+    TextAudioSpeakerCollate,
+    DistributedBucketSampler
+)
+from models import (
+    SynthesizerTrn,
+    MultiPeriodDiscriminator,
+    DurationDiscriminator,
+)
+from losses import (
+    generator_loss,
+    discriminator_loss,
+    feature_loss,
+    kl_loss
+)
+from mel_processing import mel_spectrogram_torch, spec_to_mel_torch
+from text.symbols import symbols
+
+torch.backends.cudnn.benchmark = True
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+torch.set_float32_matmul_precision('medium')
+global_step = 0
+
+
+def main():
+    """Assume Single Node Multi GPUs Training Only"""
+    assert torch.cuda.is_available(), "CPU training is not allowed."
+
+    n_gpus = torch.cuda.device_count()
+    os.environ['MASTER_ADDR'] = 'localhost'
+    os.environ['MASTER_PORT'] = '65280'
+
+    hps = utils.get_hparams()
+    if not hps.cont:
+           shutil.copy('./pretrained_models/D_0.pth','./logs/OUTPUT_MODEL/D_0.pth')
+           shutil.copy('./pretrained_models/G_0.pth','./logs/OUTPUT_MODEL/G_0.pth')
+           shutil.copy('./pretrained_models/DUR_0.pth','./logs/OUTPUT_MODEL/DUR_0.pth')
+    mp.spawn(run, nprocs=n_gpus, args=(n_gpus, hps,))
+
+
+def run(rank, n_gpus, hps):
+    global global_step
+    if rank == 0:
+        logger = utils.get_logger(hps.model_dir)
+        logger.info(hps)
+        utils.check_git_hash(hps.model_dir)
+        writer = SummaryWriter(log_dir=hps.model_dir)
+        writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
+
+    dist.init_process_group(backend=  'gloo' if os.name == 'nt' else 'nccl', init_method='env://', world_size=n_gpus, rank=rank)
+    torch.manual_seed(hps.train.seed)
+    torch.cuda.set_device(rank)
+
+    train_dataset = TextAudioSpeakerLoader(hps.data.training_files, hps.data)
+    train_sampler = DistributedBucketSampler(
+        train_dataset,
+        hps.train.batch_size,
+        [32, 300, 400, 500, 600, 700, 800, 900, 1000],
+        num_replicas=n_gpus,
+        rank=rank,
+        shuffle=True)
+    collate_fn = TextAudioSpeakerCollate()
+    train_loader = DataLoader(train_dataset, num_workers=2, shuffle=False, pin_memory=True,
+                              collate_fn=collate_fn, batch_sampler=train_sampler)
+    if rank == 0:
+        eval_dataset = TextAudioSpeakerLoader(hps.data.validation_files, hps.data)
+        eval_loader = DataLoader(eval_dataset, num_workers=0, shuffle=False,
+                                 batch_size=1, pin_memory=True,
+                                 drop_last=False, collate_fn=collate_fn)
+    if "use_noise_scaled_mas" in hps.model.keys() and hps.model.use_noise_scaled_mas == True:
+        print("Using noise scaled MAS for VITS2")
+        use_noise_scaled_mas = True
+        mas_noise_scale_initial = 0.01
+        noise_scale_delta = 2e-6
+    else:
+        print("Using normal MAS for VITS1")
+        use_noise_scaled_mas = False
+        mas_noise_scale_initial = 0.0
+        noise_scale_delta = 0.0
+    if "use_duration_discriminator" in hps.model.keys() and hps.model.use_duration_discriminator == True:
+        print("Using duration discriminator for VITS2")
+        use_duration_discriminator = True
+        net_dur_disc = DurationDiscriminator(
+         hps.model.hidden_channels, 
+         hps.model.hidden_channels, 
+         3, 
+         0.1, 
+         gin_channels=hps.model.gin_channels if hps.data.n_speakers != 0 else 0,
+         ).cuda(rank)
+    if "use_spk_conditioned_encoder" in hps.model.keys() and hps.model.use_spk_conditioned_encoder == True:
+        if hps.data.n_speakers == 0:
+            raise ValueError("n_speakers must be > 0 when using spk conditioned encoder to train multi-speaker model")
+        use_spk_conditioned_encoder = True
+    else:
+        print("Using normal encoder for VITS1")
+        use_spk_conditioned_encoder = False
+
+    net_g = SynthesizerTrn(
+        len(symbols),
+        hps.data.filter_length // 2 + 1,
+        hps.train.segment_size // hps.data.hop_length,
+        n_speakers=hps.data.n_speakers,
+        mas_noise_scale_initial = mas_noise_scale_initial,
+        noise_scale_delta = noise_scale_delta,
+        **hps.model).cuda(rank)
+
+    freeze_enc = getattr(hps.model, "freeze_enc", False)
+    if freeze_enc:
+        print("freeze encoder !!!")
+        for param in net_g.enc_p.parameters():
+            param.requires_grad = False
+
+    net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(rank)
+    optim_g = torch.optim.AdamW(
+        filter(lambda p: p.requires_grad, net_g.parameters()),
+        hps.train.learning_rate,
+        betas=hps.train.betas,
+        eps=hps.train.eps)
+    optim_d = torch.optim.AdamW(
+        net_d.parameters(),
+        hps.train.learning_rate,
+        betas=hps.train.betas,
+        eps=hps.train.eps)
+    if net_dur_disc is not None:
+        optim_dur_disc = torch.optim.AdamW(
+        net_dur_disc.parameters(),
+        hps.train.learning_rate,
+        betas=hps.train.betas,
+        eps=hps.train.eps)
+    else:
+        optim_dur_disc = None
+    net_g = DDP(net_g, device_ids=[rank], find_unused_parameters=True)
+    net_d = DDP(net_d, device_ids=[rank], find_unused_parameters=True)
+    if net_dur_disc is not None:
+        net_dur_disc = DDP(net_dur_disc, device_ids=[rank], find_unused_parameters=True)
+
+    pretrain_dir = None
+    if pretrain_dir is None:
+        try:
+            if net_dur_disc is not None:
+                _, optim_dur_disc, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "DUR_*.pth"), net_dur_disc, optim_dur_disc, skip_optimizer=not hps.cont)
+            _, optim_g, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g,
+                                                   optim_g, skip_optimizer=not hps.cont)
+            _, optim_d, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d,
+                                                   optim_d, skip_optimizer=not hps.cont)
+            
+            epoch_str = max(epoch_str, 1)
+            global_step = (epoch_str - 1) * len(train_loader)
+        except Exception as e:
+            print(e)
+            epoch_str = 1
+            global_step = 0
+    else:
+        _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(pretrain_dir, "G_*.pth"), net_g,
+                                                   optim_g, True)
+        _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(pretrain_dir, "D_*.pth"), net_d,
+                                                   optim_d, True)
+
+
+
+    scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2)
+    scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2)
+    if net_dur_disc is not None:
+        scheduler_dur_disc = torch.optim.lr_scheduler.ExponentialLR(optim_dur_disc, gamma=hps.train.lr_decay, last_epoch=epoch_str-2)
+    else:
+        scheduler_dur_disc = None
+    scaler = GradScaler(enabled=hps.train.fp16_run)
+
+    for epoch in range(epoch_str, hps.train.epochs + 1):
+        if rank == 0:
+            train_and_evaluate(rank, epoch, hps, [net_g, net_d, net_dur_disc], [optim_g, optim_d, optim_dur_disc], [scheduler_g, scheduler_d, scheduler_dur_disc], scaler, [train_loader, eval_loader], logger, [writer, writer_eval])
+        else:
+            train_and_evaluate(rank, epoch, hps, [net_g, net_d, net_dur_disc], [optim_g, optim_d, optim_dur_disc], [scheduler_g, scheduler_d, scheduler_dur_disc], scaler, [train_loader, None], None, None)
+        scheduler_g.step()
+        scheduler_d.step()
+        if net_dur_disc is not None:
+            scheduler_dur_disc.step()
+
+
+def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers):
+    net_g, net_d, net_dur_disc = nets
+    optim_g, optim_d, optim_dur_disc = optims
+    scheduler_g, scheduler_d, scheduler_dur_disc = schedulers
+    train_loader, eval_loader = loaders
+    if writers is not None:
+        writer, writer_eval = writers
+
+    train_loader.batch_sampler.set_epoch(epoch)
+    global global_step
+
+    net_g.train()
+    net_d.train()
+    if net_dur_disc is not None:
+        net_dur_disc.train()
+    for batch_idx, (x, x_lengths, spec, spec_lengths, y, y_lengths, speakers, tone, language, bert) in tqdm(enumerate(train_loader)):
+        if net_g.module.use_noise_scaled_mas:
+            current_mas_noise_scale = net_g.module.mas_noise_scale_initial - net_g.module.noise_scale_delta * global_step
+            net_g.module.current_mas_noise_scale = max(current_mas_noise_scale, 0.0)
+        x, x_lengths = x.cuda(rank, non_blocking=True), x_lengths.cuda(rank, non_blocking=True)
+        spec, spec_lengths = spec.cuda(rank, non_blocking=True), spec_lengths.cuda(rank, non_blocking=True)
+        y, y_lengths = y.cuda(rank, non_blocking=True), y_lengths.cuda(rank, non_blocking=True)
+        speakers = speakers.cuda(rank, non_blocking=True)
+        tone = tone.cuda(rank, non_blocking=True)
+        language = language.cuda(rank, non_blocking=True)
+        bert = bert.cuda(rank, non_blocking=True)
+
+        with autocast(enabled=hps.train.fp16_run):
+            y_hat, l_length, attn, ids_slice, x_mask, z_mask, \
+                (z, z_p, m_p, logs_p, m_q, logs_q), (hidden_x, logw, logw_) = net_g(x, x_lengths, spec, spec_lengths, speakers, tone, language, bert)
+            mel = spec_to_mel_torch(
+                spec,
+                hps.data.filter_length,
+                hps.data.n_mel_channels,
+                hps.data.sampling_rate,
+                hps.data.mel_fmin,
+                hps.data.mel_fmax)
+            y_mel = commons.slice_segments(mel, ids_slice, hps.train.segment_size // hps.data.hop_length)
+            y_hat_mel = mel_spectrogram_torch(
+                y_hat.squeeze(1),
+                hps.data.filter_length,
+                hps.data.n_mel_channels,
+                hps.data.sampling_rate,
+                hps.data.hop_length,
+                hps.data.win_length,
+                hps.data.mel_fmin,
+                hps.data.mel_fmax
+            )
+
+            y = commons.slice_segments(y, ids_slice * hps.data.hop_length, hps.train.segment_size)  # slice
+
+            # Discriminator
+            y_d_hat_r, y_d_hat_g, _, _ = net_d(y, y_hat.detach())
+            with autocast(enabled=False):
+                loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_d_hat_r, y_d_hat_g)
+                loss_disc_all = loss_disc
+            if net_dur_disc is not None:
+                y_dur_hat_r, y_dur_hat_g = net_dur_disc(hidden_x.detach(), x_mask.detach(), logw.detach(), logw_.detach())
+                with autocast(enabled=False):
+                 # TODO: I think need to mean using the mask, but for now, just mean all
+                    loss_dur_disc, losses_dur_disc_r, losses_dur_disc_g = discriminator_loss(y_dur_hat_r, y_dur_hat_g)
+                    loss_dur_disc_all = loss_dur_disc
+                optim_dur_disc.zero_grad()
+                scaler.scale(loss_dur_disc_all).backward()
+                scaler.unscale_(optim_dur_disc)
+                grad_norm_dur_disc = commons.clip_grad_value_(net_dur_disc.parameters(), None)
+                scaler.step(optim_dur_disc)
+
+        optim_d.zero_grad()
+        scaler.scale(loss_disc_all).backward()
+        scaler.unscale_(optim_d)
+        grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
+        scaler.step(optim_d)
+
+        with autocast(enabled=hps.train.fp16_run):
+            # Generator
+            y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
+            if net_dur_disc is not None:
+                y_dur_hat_r, y_dur_hat_g = net_dur_disc(hidden_x, x_mask, logw, logw_)
+            with autocast(enabled=False):
+                loss_dur = torch.sum(l_length.float())
+                loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
+                loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
+
+                loss_fm = feature_loss(fmap_r, fmap_g)
+                loss_gen, losses_gen = generator_loss(y_d_hat_g)
+                loss_gen_all = loss_gen + loss_fm + loss_mel + loss_dur + loss_kl
+                if net_dur_disc is not None:
+                    loss_dur_gen, losses_dur_gen = generator_loss(y_dur_hat_g)
+                    loss_gen_all += loss_dur_gen
+        optim_g.zero_grad()
+        scaler.scale(loss_gen_all).backward()
+        scaler.unscale_(optim_g)
+        grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
+        scaler.step(optim_g)
+        scaler.update()
+
+        if rank == 0:
+            if global_step % hps.train.log_interval == 0:
+                lr = optim_g.param_groups[0]['lr']
+                losses = [loss_disc, loss_gen, loss_fm, loss_mel, loss_dur, loss_kl]
+                logger.info('Train Epoch: {} [{:.0f}%]'.format(
+                    epoch,
+                    100. * batch_idx / len(train_loader)))
+                logger.info([x.item() for x in losses] + [global_step, lr])
+
+                scalar_dict = {"loss/g/total": loss_gen_all, "loss/d/total": loss_disc_all, "learning_rate": lr,
+                               "grad_norm_d": grad_norm_d, "grad_norm_g": grad_norm_g}
+                scalar_dict.update(
+                    {"loss/g/fm": loss_fm, "loss/g/mel": loss_mel, "loss/g/dur": loss_dur, "loss/g/kl": loss_kl})
+                scalar_dict.update({"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)})
+                scalar_dict.update({"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)})
+                scalar_dict.update({"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)})
+          
+                image_dict = {
+                    "slice/mel_org": utils.plot_spectrogram_to_numpy(y_mel[0].data.cpu().numpy()),
+                    "slice/mel_gen": utils.plot_spectrogram_to_numpy(y_hat_mel[0].data.cpu().numpy()),
+                    "all/mel": utils.plot_spectrogram_to_numpy(mel[0].data.cpu().numpy()),
+                    "all/attn": utils.plot_alignment_to_numpy(attn[0, 0].data.cpu().numpy())
+                }
+                utils.summarize(
+                    writer=writer,
+                    global_step=global_step,
+                    images=image_dict,
+                    scalars=scalar_dict)
+
+            if global_step % hps.train.eval_interval == 0:
+                evaluate(hps, net_g, eval_loader, writer_eval)
+                utils.save_checkpoint(net_g, optim_g, hps.train.learning_rate, epoch,
+                                      os.path.join(hps.model_dir, "G_{}.pth".format(global_step)))
+                utils.save_checkpoint(net_d, optim_d, hps.train.learning_rate, epoch,
+                                      os.path.join(hps.model_dir, "D_{}.pth".format(global_step)))
+                if net_dur_disc is not None:
+                    utils.save_checkpoint(net_dur_disc, optim_dur_disc, hps.train.learning_rate, epoch, os.path.join(hps.model_dir, "DUR_{}.pth".format(global_step)))    
+                keep_ckpts = getattr(hps.train, 'keep_ckpts', 5)
+                if keep_ckpts > 0:
+                    utils.clean_checkpoints(path_to_models=hps.model_dir, n_ckpts_to_keep=keep_ckpts, sort_by_time=True)
+
+
+        global_step += 1
+
+    if rank == 0:
+        logger.info('====> Epoch: {}'.format(epoch))
+
+
+
+def evaluate(hps, generator, eval_loader, writer_eval):
+    generator.eval()
+    image_dict = {}
+    audio_dict = {}
+    print("Evaluating ...")
+    with torch.no_grad():
+        for batch_idx, (x, x_lengths, spec, spec_lengths, y, y_lengths, speakers, tone, language, bert) in enumerate(eval_loader):
+            x, x_lengths = x.cuda(), x_lengths.cuda()
+            spec, spec_lengths = spec.cuda(), spec_lengths.cuda()
+            y, y_lengths = y.cuda(), y_lengths.cuda()
+            speakers = speakers.cuda()
+            bert = bert.cuda()
+            tone = tone.cuda()
+            language = language.cuda()
+            for use_sdp in [True, False]:
+                y_hat, attn, mask, *_ = generator.module.infer(x, x_lengths, speakers, tone, language, bert, y=spec, max_len=1000, sdp_ratio=0.0 if not use_sdp else 1.0)
+                y_hat_lengths = mask.sum([1, 2]).long() * hps.data.hop_length
+
+                mel = spec_to_mel_torch(
+                    spec,
+                    hps.data.filter_length,
+                    hps.data.n_mel_channels,
+                    hps.data.sampling_rate,
+                    hps.data.mel_fmin,
+                    hps.data.mel_fmax)
+                y_hat_mel = mel_spectrogram_torch(
+                    y_hat.squeeze(1).float(),
+                    hps.data.filter_length,
+                    hps.data.n_mel_channels,
+                    hps.data.sampling_rate,
+                    hps.data.hop_length,
+                    hps.data.win_length,
+                    hps.data.mel_fmin,
+                    hps.data.mel_fmax
+                )
+                image_dict.update({
+                    f"gen/mel_{batch_idx}": utils.plot_spectrogram_to_numpy(y_hat_mel[0].cpu().numpy())
+                })
+                audio_dict.update({
+                    f"gen/audio_{batch_idx}_{use_sdp}": y_hat[0, :, :y_hat_lengths[0]]
+                })
+                image_dict.update({f"gt/mel_{batch_idx}": utils.plot_spectrogram_to_numpy(mel[0].cpu().numpy())})
+                audio_dict.update({f"gt/audio_{batch_idx}": y[0, :, :y_lengths[0]]})
+
+    utils.summarize(
+        writer=writer_eval,
+        global_step=global_step,
+        images=image_dict,
+        audios=audio_dict,
+        audio_sampling_rate=hps.data.sampling_rate
+    )
+    generator.train()
+
+if __name__ == "__main__":
+    main()

transcribe_genshin.py

@@ -0,0 +1,78 @@
+# coding=gbk
+import os
+import argparse
+import librosa
+import numpy as np
+from multiprocessing import Pool, cpu_count
+
+import soundfile
+from scipy.io import wavfile
+from tqdm import tqdm
+
+global speaker_annos
+speaker_annos = []
+
+def process(item):  
+    spkdir, wav_name, args = item
+    speaker = spkdir.replace("\\", "/").split("/")[-1]
+    wav_path = os.path.join(args.in_dir, speaker, wav_name)
+    if os.path.exists(wav_path) and '.wav' in wav_path:
+        os.makedirs(os.path.join(args.out_dir, speaker), exist_ok=True)
+        wav, sr = librosa.load(wav_path, sr=args.sr)
+        soundfile.write(
+            os.path.join(args.out_dir, speaker, wav_name),
+            wav,
+            sr
+        )
+
+def process_text(item):
+    spkdir, wav_name, args = item
+    speaker = spkdir.replace("\\", "/").split("/")[-1]
+    wav_path = os.path.join(args.in_dir, speaker, wav_name)
+    global speaker_annos
+    tr_name = wav_name.replace('.wav', '')
+    with open(args.out_dir+'/'+speaker+'/'+tr_name+'.lab', "r", encoding="utf-8") as file:
+             text = file.read()
+    text = text.replace("{NICKNAME}",'������')
+    text = text.replace("{M#��}{F#��}",'��')
+    text = text.replace("{M#��}{F#��}",'��')
+    substring = "{M#����}{F#���}"  
+    if substring in text:
+        if tr_name.endswith("a"):
+           text = text.replace("{M#����}{F#���}",'����')
+        if tr_name.endswith("b"):
+           text = text.replace("{M#����}{F#���}",'���')
+    text = text.replace("#",'')   
+    text = "ZH|" + text + "\n" #
+    speaker_annos.append(args.out_dir+'/'+speaker+'/'+wav_name+ "|" + speaker + "|" + text)
+
+
+
+if __name__ == "__main__":
+    parent_dir = "./genshin_dataset/"
+    speaker_names = list(os.walk(parent_dir))[0][1]
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--sr", type=int, default=44100, help="sampling rate")
+    parser.add_argument("--in_dir", type=str, default="./genshin_dataset", help="path to source dir")
+    parser.add_argument("--out_dir", type=str, default="./genshin_dataset", help="path to target dir")
+    args = parser.parse_args()
+    # processs = 8
+    processs = cpu_count()-2 if cpu_count() >4 else 1
+    pool = Pool(processes=processs)
+
+    for speaker in os.listdir(args.in_dir):
+        spk_dir = os.path.join(args.in_dir, speaker)
+        if os.path.isdir(spk_dir):
+            print(spk_dir)
+            for _ in tqdm(pool.imap_unordered(process, [(spk_dir, i, args) for i in os.listdir(spk_dir) if i.endswith("wav")])):
+                pass
+            for i in os.listdir(spk_dir):
+               if i.endswith("wav"):
+                  pro=(spk_dir, i, args)
+                  process_text(pro)
+    if len(speaker_annos) == 0:
+        print("transcribe error!!!")
+    with open("./filelists/short_character_anno.list", 'w', encoding='utf-8') as f:
+        for line in speaker_annos:
+            f.write(line)
+    print("transcript file finished.")