Duplicate from kevinwang676/Bark-Voice-Cloning

Co-authored-by: Kevin Wang <kevinwang676@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,34 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,14 @@
+__pycache__/
+/outputs
+/speakers
+.vs
+*.npz
+*.wav
+*.npy
+.vs/
+/models
+/bark_ui_enhanced.egg-info
+/build/lib/bark
+*.pth
+*.pt
+*.zip

Dockerfile

@@ -0,0 +1,38 @@
+FROM debian:stable
+
+# Install system packages
+RUN apt update && apt install -y git pip
+
+# Create non-root user
+RUN useradd -m -d /bark bark
+
+# Run as new user
+USER bark
+WORKDIR /bark
+
+# Clone git repo
+RUN git clone https://github.com/C0untFloyd/bark-gui 
+
+# Switch to git directory
+WORKDIR /bark/bark-gui
+
+# Append pip bin path to PATH
+ENV PATH=$PATH:/bark/.local/bin
+
+# Install dependancies
+RUN pip install .
+RUN pip install -r requirements.txt
+
+# List on all addresses, since we are in a container.
+RUN sed -i "s/server_name: ''/server_name: 0.0.0.0/g" ./config.yaml
+
+# Suggested volumes
+VOLUME /bark/bark-gui/assets/prompts/custom
+VOLUME /bark/bark-gui/models
+VOLUME /bark/.cache/huggingface/hub
+
+# Default port for web-ui
+EXPOSE 7860/tcp
+
+# Start script
+CMD python3 webui.py

README.md

@@ -0,0 +1,14 @@
+---
+title: Bark Voice Cloning
+emoji: 🎶
+colorFrom: yellow
+colorTo: pink
+sdk: gradio
+sdk_version: 3.34.0
+app_file: app.py
+pinned: false
+license: mit
+duplicated_from: kevinwang676/Bark-Voice-Cloning
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,401 @@
+from cProfile import label
+import dataclasses
+from distutils.command.check import check
+from doctest import Example
+import gradio as gr
+import os
+import sys
+import numpy as np
+import logging
+import torch
+import pytorch_seed
+import time
+
+from xml.sax import saxutils
+from bark.api import generate_with_settings
+from bark.api import save_as_prompt
+from util.settings import Settings
+#import nltk
+
+
+from bark import SAMPLE_RATE
+from cloning.clonevoice import clone_voice
+from bark.generation import SAMPLE_RATE, preload_models, _load_history_prompt, codec_decode
+from scipy.io.wavfile import write as write_wav
+from util.parseinput import split_and_recombine_text, build_ssml, is_ssml, create_clips_from_ssml
+from datetime import datetime
+from tqdm.auto import tqdm
+from util.helper import create_filename, add_id3_tag
+from swap_voice import swap_voice_from_audio
+from training.training_prepare import prepare_semantics_from_text, prepare_wavs_from_semantics
+from training.train import training_prepare_files, train
+
+settings = Settings('config.yaml')
+
+
+def generate_text_to_speech(text, selected_speaker, text_temp, waveform_temp, eos_prob, quick_generation, complete_settings, seed, batchcount, progress=gr.Progress(track_tqdm=True)):
+    # Chunk the text into smaller pieces then combine the generated audio
+
+    # generation settings
+    if selected_speaker == 'None':
+        selected_speaker = None
+
+    voice_name = selected_speaker
+
+    if text == None or len(text) < 1:
+       if selected_speaker == None:
+            raise gr.Error('No text entered!')
+
+       # Extract audio data from speaker if no text and speaker selected
+       voicedata = _load_history_prompt(voice_name)
+       audio_arr = codec_decode(voicedata["fine_prompt"])
+       result = create_filename(settings.output_folder_path, "None", "extract",".wav")
+       save_wav(audio_arr, result)
+       return result
+
+    if batchcount < 1:
+        batchcount = 1
+
+
+    silenceshort = np.zeros(int((float(settings.silence_sentence) / 1000.0) * SAMPLE_RATE), dtype=np.int16)  # quarter second of silence
+    silencelong = np.zeros(int((float(settings.silence_speakers) / 1000.0) * SAMPLE_RATE), dtype=np.float32)  # half a second of silence
+    use_last_generation_as_history = "Use last generation as history" in complete_settings
+    save_last_generation = "Save generation as Voice" in complete_settings
+    for l in range(batchcount):
+        currentseed = seed
+        if seed != None and seed > 2**32 - 1:
+            logger.warning(f"Seed {seed} > 2**32 - 1 (max), setting to random")
+            currentseed = None
+        if currentseed == None or currentseed <= 0:
+            currentseed = np.random.default_rng().integers(1, 2**32 - 1)
+        assert(0 < currentseed and currentseed < 2**32)
+
+        progress(0, desc="Generating")
+
+        full_generation = None
+
+        all_parts = []
+        complete_text = ""
+        text = text.lstrip()
+        if is_ssml(text):
+            list_speak = create_clips_from_ssml(text)
+            prev_speaker = None
+            for i, clip in tqdm(enumerate(list_speak), total=len(list_speak)):
+                selected_speaker = clip[0]
+                # Add pause break between speakers
+                if i > 0 and selected_speaker != prev_speaker:
+                    all_parts += [silencelong.copy()]
+                prev_speaker = selected_speaker
+                text = clip[1]
+                text = saxutils.unescape(text)
+                if selected_speaker == "None":
+                    selected_speaker = None
+
+                print(f"\nGenerating Text ({i+1}/{len(list_speak)}) -> {selected_speaker} (Seed {currentseed}):`{text}`")
+                complete_text += text
+                with pytorch_seed.SavedRNG(currentseed):
+                    audio_array = generate_with_settings(text_prompt=text, voice_name=selected_speaker, semantic_temp=text_temp, coarse_temp=waveform_temp, eos_p=eos_prob)
+                    currentseed = torch.random.initial_seed()
+                if len(list_speak) > 1:
+                    filename = create_filename(settings.output_folder_path, currentseed, "audioclip",".wav")
+                    save_wav(audio_array, filename)
+                    add_id3_tag(filename, text, selected_speaker, currentseed)
+
+                all_parts += [audio_array]
+        else:
+            texts = split_and_recombine_text(text, settings.input_text_desired_length, settings.input_text_max_length)
+            for i, text in tqdm(enumerate(texts), total=len(texts)):
+                print(f"\nGenerating Text ({i+1}/{len(texts)}) -> {selected_speaker} (Seed {currentseed}):`{text}`")
+                complete_text += text
+                if quick_generation == True:
+                    with pytorch_seed.SavedRNG(currentseed):
+                        audio_array = generate_with_settings(text_prompt=text, voice_name=selected_speaker, semantic_temp=text_temp, coarse_temp=waveform_temp, eos_p=eos_prob)
+                        currentseed = torch.random.initial_seed()
+                else:
+                    full_output = use_last_generation_as_history or save_last_generation
+                    if full_output:
+                        full_generation, audio_array = generate_with_settings(text_prompt=text, voice_name=voice_name, semantic_temp=text_temp, coarse_temp=waveform_temp, eos_p=eos_prob, output_full=True)
+                    else:
+                        audio_array = generate_with_settings(text_prompt=text, voice_name=voice_name, semantic_temp=text_temp, coarse_temp=waveform_temp, eos_p=eos_prob)
+
+                # Noticed this in the HF Demo - convert to 16bit int -32767/32767 - most used audio format  
+                # audio_array = (audio_array * 32767).astype(np.int16)
+
+                if len(texts) > 1:
+                    filename = create_filename(settings.output_folder_path, currentseed, "audioclip",".wav")
+                    save_wav(audio_array, filename)
+                    add_id3_tag(filename, text, selected_speaker, currentseed)
+
+                if quick_generation == False and (save_last_generation == True or use_last_generation_as_history == True):
+                    # save to npz
+                    voice_name = create_filename(settings.output_folder_path, seed, "audioclip", ".npz")
+                    save_as_prompt(voice_name, full_generation)
+                    if use_last_generation_as_history:
+                        selected_speaker = voice_name
+
+                all_parts += [audio_array]
+                # Add short pause between sentences
+                if text[-1] in "!?.\n" and i > 1:
+                    all_parts += [silenceshort.copy()]
+
+        # save & play audio
+        result = create_filename(settings.output_folder_path, currentseed, "final",".wav")
+        save_wav(np.concatenate(all_parts), result)
+        # write id3 tag with text truncated to 60 chars, as a precaution...
+        add_id3_tag(result, complete_text, selected_speaker, currentseed)
+
+    return result
+
+
+
+def save_wav(audio_array, filename):
+    write_wav(filename, SAMPLE_RATE, audio_array)
+
+def save_voice(filename, semantic_prompt, coarse_prompt, fine_prompt):
+    np.savez_compressed(
+        filename,
+        semantic_prompt=semantic_prompt,
+        coarse_prompt=coarse_prompt,
+        fine_prompt=fine_prompt
+    )
+    
+
+def on_quick_gen_changed(checkbox):
+    if checkbox == False:
+        return gr.CheckboxGroup.update(visible=True)
+    return gr.CheckboxGroup.update(visible=False)
+
+def delete_output_files(checkbox_state):
+    if checkbox_state:
+        outputs_folder = os.path.join(os.getcwd(), settings.output_folder_path)
+        if os.path.exists(outputs_folder):
+            purgedir(outputs_folder)
+    return False
+
+
+# https://stackoverflow.com/a/54494779
+def purgedir(parent):
+    for root, dirs, files in os.walk(parent):                                      
+        for item in files:
+            # Delete subordinate files                                                 
+            filespec = os.path.join(root, item)
+            os.unlink(filespec)
+        for item in dirs:
+            # Recursively perform this operation for subordinate directories   
+            purgedir(os.path.join(root, item))
+
+def convert_text_to_ssml(text, selected_speaker):
+    return build_ssml(text, selected_speaker)
+
+
+def training_prepare(selected_step, num_text_generations, progress=gr.Progress(track_tqdm=True)):
+    if selected_step == prepare_training_list[0]:
+        prepare_semantics_from_text()
+    else:
+        prepare_wavs_from_semantics()
+    return None
+
+
+def start_training(save_model_epoch, max_epochs, progress=gr.Progress(track_tqdm=True)):
+    training_prepare_files("./training/data/", "./training/data/checkpoint/hubert_base_ls960.pt")
+    train("./training/data/", save_model_epoch, max_epochs)
+    return None
+
+
+
+def apply_settings(themes, input_server_name, input_server_port, input_server_public, input_desired_len, input_max_len, input_silence_break, input_silence_speaker):
+    settings.selected_theme = themes
+    settings.server_name = input_server_name
+    settings.server_port = input_server_port
+    settings.server_share = input_server_public
+    settings.input_text_desired_length = input_desired_len
+    settings.input_text_max_length = input_max_len
+    settings.silence_sentence = input_silence_break
+    settings.silence_speaker = input_silence_speaker
+    settings.save()
+
+def restart():
+    global restart_server
+    restart_server = True
+
+
+def create_version_html():
+    python_version = ".".join([str(x) for x in sys.version_info[0:3]])
+    versions_html = f"""
+python: <span title="{sys.version}">{python_version}</span>
+ • 
+torch: {getattr(torch, '__long_version__',torch.__version__)}
+ • 
+gradio: {gr.__version__}
+"""
+    return versions_html
+
+    
+
+logger = logging.getLogger(__name__)
+APPTITLE = "Bark Voice Cloning UI"
+
+
+autolaunch = False
+
+if len(sys.argv) > 1:
+    autolaunch = "-autolaunch" in sys.argv
+
+
+if torch.cuda.is_available() == False:
+    os.environ['BARK_FORCE_CPU'] = 'True'
+    logger.warning("No CUDA detected, fallback to CPU!")
+
+print(f'smallmodels={os.environ.get("SUNO_USE_SMALL_MODELS", False)}')
+print(f'enablemps={os.environ.get("SUNO_ENABLE_MPS", False)}')
+print(f'offloadcpu={os.environ.get("SUNO_OFFLOAD_CPU", False)}')
+print(f'forcecpu={os.environ.get("BARK_FORCE_CPU", False)}')
+print(f'autolaunch={autolaunch}\n\n')
+
+#print("Updating nltk\n")
+#nltk.download('punkt')
+
+print("Preloading Models\n")
+preload_models()
+
+available_themes = ["Default", "gradio/glass", "gradio/monochrome", "gradio/seafoam", "gradio/soft", "gstaff/xkcd", "freddyaboulton/dracula_revamped", "ysharma/steampunk"]
+tokenizer_language_list = ["de","en", "pl"]
+prepare_training_list = ["Step 1: Semantics from Text","Step 2: WAV from Semantics"]
+
+seed = -1
+server_name = settings.server_name
+if len(server_name) < 1:
+    server_name = None
+server_port = settings.server_port
+if server_port <= 0:
+    server_port = None
+global run_server
+global restart_server
+
+run_server = True
+
+while run_server:
+    # Collect all existing speakers/voices in dir
+    speakers_list = []
+
+    for root, dirs, files in os.walk("./bark/assets/prompts"):
+        for file in files:
+            if file.endswith(".npz"):
+                pathpart = root.replace("./bark/assets/prompts", "")
+                name = os.path.join(pathpart, file[:-4])
+                if name.startswith("/") or name.startswith("\\"):
+                     name = name[1:]
+                speakers_list.append(name)
+
+    speakers_list = sorted(speakers_list, key=lambda x: x.lower())
+    speakers_list.insert(0, 'None')
+
+    print(f'Launching {APPTITLE} Server')
+
+    # Create Gradio Blocks
+
+    with gr.Blocks(title=f"{APPTITLE}", mode=f"{APPTITLE}", theme=settings.selected_theme) as barkgui:
+        gr.Markdown("# <center>🐶🎶⭐ - Bark Voice Cloning</center>")
+        gr.Markdown("### <center>🤗 - If you like this space, please star my [github repo](https://github.com/KevinWang676/Bark-Voice-Cloning)</center>")
+        gr.Markdown("### <center>🎡 - Based on [bark-gui](https://github.com/C0untFloyd/bark-gui)</center>")
+        gr.Markdown(f""" You can duplicate and use it with a GPU: <a href="https://huggingface.co/spaces/{os.getenv('SPACE_ID')}?duplicate=true"><img style="display: inline; margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space" /></a>
+                         or open in [Colab](https://colab.research.google.com/github/KevinWang676/Bark-Voice-Cloning/blob/main/Bark_Voice_Cloning_UI.ipynb) for quick start 🌟
+                    """)
+
+        with gr.Tab("🎙️ - Clone Voice"):
+            with gr.Row():
+                input_audio_filename = gr.Audio(label="Input audio.wav", source="upload", type="filepath")
+            #transcription_text = gr.Textbox(label="Transcription Text", lines=1, placeholder="Enter Text of your Audio Sample here...")
+            with gr.Row():
+                with gr.Column():
+                    initialname = "/home/user/app/bark/assets/prompts/file"
+                    output_voice = gr.Textbox(label="Filename of trained Voice (do not change the initial name)", lines=1, placeholder=initialname, value=initialname, visible=False)
+                with gr.Column():
+                    tokenizerlang = gr.Dropdown(tokenizer_language_list, label="Base Language Tokenizer", value=tokenizer_language_list[1], visible=False)
+            with gr.Row():
+                clone_voice_button = gr.Button("Create Voice", variant="primary")
+            with gr.Row():
+                dummy = gr.Text(label="Progress")
+                npz_file = gr.File(label=".npz file")
+            speakers_list.insert(0, npz_file) # add prompt
+
+        with gr.Tab("🎵 - TTS"):
+            with gr.Row():
+                with gr.Column():
+                    placeholder = "Enter text here."
+                    input_text = gr.Textbox(label="Input Text", lines=4, placeholder=placeholder)
+                    convert_to_ssml_button = gr.Button("Convert Input Text to SSML")
+                with gr.Column():
+                        seedcomponent = gr.Number(label="Seed (default -1 = Random)", precision=0, value=-1)
+                        batchcount = gr.Number(label="Batch count", precision=0, value=1)
+           
+            with gr.Row():
+                with gr.Column():
+                    gr.Markdown("[Voice Prompt Library](https://suno-ai.notion.site/8b8e8749ed514b0cbf3f699013548683?v=bc67cff786b04b50b3ceb756fd05f68c)")
+                    speaker = gr.Dropdown(speakers_list, value=speakers_list[0], label="Voice (Choose “file” if you wanna use the custom voice)")
+                    
+                with gr.Column():
+                    text_temp = gr.Slider(0.1, 1.0, value=0.6, label="Generation Temperature", info="1.0 more diverse, 0.1 more conservative")
+                    waveform_temp = gr.Slider(0.1, 1.0, value=0.7, label="Waveform temperature", info="1.0 more diverse, 0.1 more conservative")
+
+            with gr.Row():
+                with gr.Column():
+                    quick_gen_checkbox = gr.Checkbox(label="Quick Generation", value=True)
+                    settings_checkboxes = ["Use last generation as history", "Save generation as Voice"]
+                    complete_settings = gr.CheckboxGroup(choices=settings_checkboxes, value=settings_checkboxes, label="Detailed Generation Settings", type="value", interactive=True, visible=False)
+                with gr.Column():
+                    eos_prob = gr.Slider(0.0, 0.5, value=0.05, label="End of sentence probability")
+
+            with gr.Row():
+                with gr.Column():
+                    tts_create_button = gr.Button("Generate", variant="primary")
+                with gr.Column():
+                    hidden_checkbox = gr.Checkbox(visible=False)
+                    button_stop_generation = gr.Button("Stop generation")
+            with gr.Row():
+                output_audio = gr.Audio(label="Generated Audio", type="filepath")
+
+        with gr.Tab("🔮 - Voice Conversion"):
+            with gr.Row():
+                 swap_audio_filename = gr.Audio(label="Input audio.wav to swap voice", source="upload", type="filepath")
+            with gr.Row():
+                 with gr.Column():
+                     swap_tokenizer_lang = gr.Dropdown(tokenizer_language_list, label="Base Language Tokenizer", value=tokenizer_language_list[1])
+                     swap_seed = gr.Number(label="Seed (default -1 = Random)", precision=0, value=-1)
+                 with gr.Column():
+                     speaker_swap = gr.Dropdown(speakers_list, value=speakers_list[0], label="Voice (Choose “file” if you wanna use the custom voice)")
+                     swap_batchcount = gr.Number(label="Batch count", precision=0, value=1)
+            with gr.Row():
+                swap_voice_button = gr.Button("Generate", variant="primary")
+            with gr.Row():
+                output_swap = gr.Audio(label="Generated Audio", type="filepath")
+
+   
+        quick_gen_checkbox.change(fn=on_quick_gen_changed, inputs=quick_gen_checkbox, outputs=complete_settings)
+        convert_to_ssml_button.click(convert_text_to_ssml, inputs=[input_text, speaker],outputs=input_text)
+        gen_click = tts_create_button.click(generate_text_to_speech, inputs=[input_text, speaker, text_temp, waveform_temp, eos_prob, quick_gen_checkbox, complete_settings, seedcomponent, batchcount],outputs=output_audio)
+        button_stop_generation.click(fn=None, inputs=None, outputs=None, cancels=[gen_click])
+        
+
+
+        swap_voice_button.click(swap_voice_from_audio, inputs=[swap_audio_filename, speaker_swap, swap_tokenizer_lang, swap_seed, swap_batchcount], outputs=output_swap)
+        clone_voice_button.click(clone_voice, inputs=[input_audio_filename, output_voice], outputs=[dummy, npz_file])
+
+
+        restart_server = False
+        try:
+            barkgui.queue().launch(show_error=True)
+        except:
+            restart_server = True
+            run_server = False
+        try:
+            while restart_server == False:
+                time.sleep(1.0)
+        except (KeyboardInterrupt, OSError):
+            print("Keyboard interruption in main thread... closing server.")
+            run_server = False
+        barkgui.close()
+
+
+
+

bark/__init__.py

@@ -0,0 +1,2 @@
+from .api import generate_audio, text_to_semantic, semantic_to_waveform, save_as_prompt
+from .generation import SAMPLE_RATE, preload_models

bark/api.py

@@ -0,0 +1,158 @@
+from typing import Dict, Optional, Union
+
+import numpy as np
+
+from .generation import codec_decode, generate_coarse, generate_fine, generate_text_semantic
+
+
+def generate_with_settings(text_prompt, semantic_temp=0.6, eos_p=0.2, coarse_temp=0.7, fine_temp=0.5, voice_name=None, output_full=False):
+
+    # generation with more control
+    x_semantic = generate_text_semantic(
+        text_prompt,
+        history_prompt=voice_name,
+        temp=semantic_temp,
+        min_eos_p = eos_p,
+        use_kv_caching=True
+    )
+
+    x_coarse_gen = generate_coarse(
+        x_semantic,
+        history_prompt=voice_name,
+        temp=coarse_temp,
+        use_kv_caching=True
+    )
+    x_fine_gen = generate_fine(
+        x_coarse_gen,
+        history_prompt=voice_name,
+        temp=fine_temp,
+    )
+
+    if output_full:
+        full_generation = {
+            'semantic_prompt': x_semantic,
+            'coarse_prompt': x_coarse_gen,
+            'fine_prompt': x_fine_gen
+        }
+        return full_generation, codec_decode(x_fine_gen)
+    return codec_decode(x_fine_gen)
+
+
+def text_to_semantic(
+    text: str,
+    history_prompt: Optional[Union[Dict, str]] = None,
+    temp: float = 0.7,
+    silent: bool = False,
+):
+    """Generate semantic array from text.
+
+    Args:
+        text: text to be turned into audio
+        history_prompt: history choice for audio cloning
+        temp: generation temperature (1.0 more diverse, 0.0 more conservative)
+        silent: disable progress bar
+
+    Returns:
+        numpy semantic array to be fed into `semantic_to_waveform`
+    """
+    x_semantic = generate_text_semantic(
+        text,
+        history_prompt=history_prompt,
+        temp=temp,
+        silent=silent,
+        use_kv_caching=True
+    )
+    return x_semantic
+
+
+def semantic_to_waveform(
+    semantic_tokens: np.ndarray,
+    history_prompt: Optional[Union[Dict, str]] = None,
+    temp: float = 0.7,
+    silent: bool = False,
+    output_full: bool = False,
+):
+    """Generate audio array from semantic input.
+
+    Args:
+        semantic_tokens: semantic token output from `text_to_semantic`
+        history_prompt: history choice for audio cloning
+        temp: generation temperature (1.0 more diverse, 0.0 more conservative)
+        silent: disable progress bar
+        output_full: return full generation to be used as a history prompt
+
+    Returns:
+        numpy audio array at sample frequency 24khz
+    """
+    coarse_tokens = generate_coarse(
+        semantic_tokens,
+        history_prompt=history_prompt,
+        temp=temp,
+        silent=silent,
+        use_kv_caching=True
+    )
+    fine_tokens = generate_fine(
+        coarse_tokens,
+        history_prompt=history_prompt,
+        temp=0.5,
+    )
+    audio_arr = codec_decode(fine_tokens)
+    if output_full:
+        full_generation = {
+            "semantic_prompt": semantic_tokens,
+            "coarse_prompt": coarse_tokens,
+            "fine_prompt": fine_tokens,
+        }
+        return full_generation, audio_arr
+    return audio_arr
+
+
+def save_as_prompt(filepath, full_generation):
+    assert(filepath.endswith(".npz"))
+    assert(isinstance(full_generation, dict))
+    assert("semantic_prompt" in full_generation)
+    assert("coarse_prompt" in full_generation)
+    assert("fine_prompt" in full_generation)
+    np.savez(filepath, **full_generation)
+
+
+def generate_audio(
+    text: str,
+    history_prompt: Optional[Union[Dict, str]] = None,
+    text_temp: float = 0.7,
+    waveform_temp: float = 0.7,
+    silent: bool = False,
+    output_full: bool = False,
+):
+    """Generate audio array from input text.
+
+    Args:
+        text: text to be turned into audio
+        history_prompt: history choice for audio cloning
+        text_temp: generation temperature (1.0 more diverse, 0.0 more conservative)
+        waveform_temp: generation temperature (1.0 more diverse, 0.0 more conservative)
+        silent: disable progress bar
+        output_full: return full generation to be used as a history prompt
+
+    Returns:
+        numpy audio array at sample frequency 24khz
+    """
+    semantic_tokens = text_to_semantic(
+        text,
+        history_prompt=history_prompt,
+        temp=text_temp,
+        silent=silent,
+    )
+    out = semantic_to_waveform(
+        semantic_tokens,
+        history_prompt=history_prompt,
+        temp=waveform_temp,
+        silent=silent,
+        output_full=output_full,
+    )
+    if output_full:
+        full_generation, audio_arr = out
+        return full_generation, audio_arr
+    else:
+        audio_arr = out
+    return audio_arr

bark/assets/prompts/announcer.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26f2d1a9e3b6fe453cf5fc8191de26cbfae6276c5b0f7c376c6a0f3c35867f83
+size 16794

bark/assets/prompts/v2/en_speaker_0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:932f40d879ba8659f1ca26319ba64ea3b0647b2050fe24313bf42b0dff1fe241
+size 28100

bark/assets/prompts/v2/en_speaker_1.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e7f18015e1ab9b6302ded1e28a971af5306a72f193bb6c411f1948a083c8578
+size 25220

bark/assets/prompts/v2/en_speaker_2.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d218990680ece5f2d4fc18ea4783b016b3ae353ec413eaee2058f2d57263c9b3
+size 26236

bark/assets/prompts/v2/en_speaker_3.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92c2e2a29145c83738e9b63f082fd1c873d9422468a155463cb27f814aeaea66
+size 34980

bark/assets/prompts/v2/en_speaker_4.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:992f91991a9a5359d72f00b09a11a550e71bb8ebfc0cfd877e39d7d41f98b714
+size 23780

bark/assets/prompts/v2/en_speaker_5.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18831c3f6014e4a2ff60ad5169b1fae06e28ed07f43f8a3616aafb84515091bf
+size 24740

bark/assets/prompts/v2/en_speaker_6.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fab38dc6b6bc9226bcc414f4c5a9524bc1b2441865a586153fb620127a8faa4e
+size 25540

bark/assets/prompts/v2/en_speaker_7.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f4c4eb33f5994be8de5cfd1744ebce13da1618a6da3a7d244514178c61ef7db
+size 22716

bark/assets/prompts/v2/en_speaker_8.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8fc9f11b539588f51bbf78150a73e0365c49b2306bd72e5a22b28ef09c4fb15d
+size 23300

bark/assets/prompts/v2/en_speaker_9.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78b3ba32eb9aeb9ed34556856c40633ecc8332d1c3ae3c81e6f5015ac3eefbd5
+size 30180

bark/assets/prompts/v2/zh_speaker_0.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd7ac118a3e944b3f20c89f2446056a00850a630ee16318922acc6572ce80929
+size 20636

bark/assets/prompts/v2/zh_speaker_1.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0eacf5c862dfd3c5ac825f2ebb26f323e64309cb712e7e264cbd31c5bca3f038
+size 19836

bark/assets/prompts/v2/zh_speaker_2.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e324b47f8250e5798c314f395d4e049575e7ca369d0b6074e91c7bba70e9f26d
+size 21060

bark/assets/prompts/v2/zh_speaker_3.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98c476abc7bf634ffb2d71d363284e7bd8c8abd5e33ec5ca21d4aa5b15730d18
+size 31300

bark/assets/prompts/v2/zh_speaker_4.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1fa8673a9895ad3302d13ac94193b5ad5da481f1cc276e6181fa895acaae133b
+size 29964

bark/assets/prompts/v2/zh_speaker_5.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:226edfe5fabc72eeb83a13e350599bc8babe5adc2264b3cdb661fd1258dc4044
+size 17436

bark/assets/prompts/v2/zh_speaker_6.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:285d51fbe81cc263636b5b487fbb6633e6f3cf92c53ca9ab8e6b7f55d4b4a31d
+size 16900

bark/assets/prompts/v2/zh_speaker_7.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0967cdb14ffa79895747b0d52df9f15bdad80d6c55b7630894345c9a7ec87c91
+size 21060

bark/assets/prompts/v2/zh_speaker_8.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c028f78530013f29ab8c0c1cf4fe2138106fbe5252951f5f36e0168056779549
+size 19300

bark/assets/prompts/v2/zh_speaker_9.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6265bb827008d7af8a45a8e057fe3e91efb347d56208180a9ed990ad54e4d75e
+size 16156

bark/generation.py

@@ -0,0 +1,864 @@
+import contextlib
+import gc
+import os
+import re
+import requests
+import gc
+import sys
+
+from encodec import EncodecModel
+import funcy
+import logging
+import numpy as np
+from scipy.special import softmax
+import torch
+import torch.nn.functional as F
+import tqdm
+from transformers import BertTokenizer
+from huggingface_hub import hf_hub_download, hf_hub_url
+
+from .model import GPTConfig, GPT
+from .model_fine import FineGPT, FineGPTConfig
+from .settings import initenv
+
+initenv(sys.argv)
+global_force_cpu = os.environ.get("BARK_FORCE_CPU", False)
+if (
+    global_force_cpu != True and
+    torch.cuda.is_available() and
+    hasattr(torch.cuda, "amp") and
+    hasattr(torch.cuda.amp, "autocast") and
+    hasattr(torch.cuda, "is_bf16_supported") and
+    torch.cuda.is_bf16_supported()
+):
+    autocast = funcy.partial(torch.cuda.amp.autocast, dtype=torch.bfloat16)
+else:
+    @contextlib.contextmanager
+    def autocast():
+        yield
+
+
+# hold models in global scope to lazy load
+global models
+models = {}
+
+global models_devices
+models_devices = {}
+
+
+CONTEXT_WINDOW_SIZE = 1024
+
+SEMANTIC_RATE_HZ = 49.9
+SEMANTIC_VOCAB_SIZE = 10_000
+
+CODEBOOK_SIZE = 1024
+N_COARSE_CODEBOOKS = 2
+N_FINE_CODEBOOKS = 8
+COARSE_RATE_HZ = 75
+
+SAMPLE_RATE = 24_000
+
+
+SUPPORTED_LANGS = [
+    ("English", "en"),
+    ("German", "de"),
+    ("Spanish", "es"),
+    ("French", "fr"),
+    ("Hindi", "hi"),
+    ("Italian", "it"),
+    ("Japanese", "ja"),
+    ("Korean", "ko"),
+    ("Polish", "pl"),
+    ("Portuguese", "pt"),
+    ("Russian", "ru"),
+    ("Turkish", "tr"),
+    ("Chinese", "zh"),
+]
+
+ALLOWED_PROMPTS = {"announcer"}
+for _, lang in SUPPORTED_LANGS:
+    for prefix in ("", f"v2{os.path.sep}"):
+        for n in range(10):
+            ALLOWED_PROMPTS.add(f"{prefix}{lang}_speaker_{n}")
+
+
+logger = logging.getLogger(__name__)
+
+
+CUR_PATH = os.path.dirname(os.path.abspath(__file__))
+
+
+#default_cache_dir = os.path.join(os.path.expanduser("~"), ".cache")
+#CACHE_DIR = os.path.join(os.getenv("XDG_CACHE_HOME", default_cache_dir), "suno", "bark_v0")
+#CACHE_DIR = os.path.join(os.getcwd(), "models"
+CACHE_DIR = "./models"
+
+
+def _cast_bool_env_var(s):
+    return s.lower() in ('true', '1', 't')
+
+USE_SMALL_MODELS = _cast_bool_env_var(os.environ.get("SUNO_USE_SMALL_MODELS", "False"))
+GLOBAL_ENABLE_MPS = _cast_bool_env_var(os.environ.get("SUNO_ENABLE_MPS", "False"))
+OFFLOAD_CPU = _cast_bool_env_var(os.environ.get("SUNO_OFFLOAD_CPU", "False"))
+
+REMOTE_MODEL_PATHS = {
+    "text_small": {
+        "repo_id": "suno/bark",
+        "file_name": "text.pt",
+    },
+    "coarse_small": {
+        "repo_id": "suno/bark",
+        "file_name": "coarse.pt",
+    },
+    "fine_small": {
+        "repo_id": "suno/bark",
+        "file_name": "fine.pt",
+    },
+    "text": {
+        "repo_id": "suno/bark",
+        "file_name": "text_2.pt",
+    },
+    "coarse": {
+        "repo_id": "suno/bark",
+        "file_name": "coarse_2.pt",
+    },
+    "fine": {
+        "repo_id": "suno/bark",
+        "file_name": "fine_2.pt",
+    },
+}
+
+
+if not hasattr(torch.nn.functional, 'scaled_dot_product_attention') and torch.cuda.is_available():
+    logger.warning(
+        "torch version does not support flash attention. You will get faster" +
+        " inference speed by upgrade torch to newest nightly version."
+    )
+
+
+def grab_best_device(use_gpu=True):
+    if torch.cuda.device_count() > 0 and use_gpu:
+        device = "cuda"
+    elif torch.backends.mps.is_available() and use_gpu and GLOBAL_ENABLE_MPS:
+        device = "mps"
+    else:
+        device = "cpu"
+    return device
+
+
+def _get_ckpt_path(model_type, use_small=False):
+    key = model_type
+    if use_small or USE_SMALL_MODELS:
+        key += "_small"
+    return os.path.join(CACHE_DIR, REMOTE_MODEL_PATHS[key]["file_name"])
+
+"""
+def _download(from_hf_path, file_name, destfilename):
+    os.makedirs(CACHE_DIR, exist_ok=True)
+    hf_hub_download(repo_id=from_hf_path, filename=file_name, local_dir=CACHE_DIR, local_dir_use_symlinks=False)
+    # Bug in original repo? Downloaded name differs from expected...
+    if not os.path.exists(destfilename):
+        localname = os.path.join(CACHE_DIR, file_name)
+        os.rename(localname, destfilename)
+"""
+def _download(from_hf_path, file_name):
+    os.makedirs(CACHE_DIR, exist_ok=True)
+    hf_hub_download(repo_id=from_hf_path, filename=file_name, local_dir=CACHE_DIR)
+
+
+class InferenceContext:
+    def __init__(self, benchmark=False):
+        # we can't expect inputs to be the same length, so disable benchmarking by default
+        self._chosen_cudnn_benchmark = benchmark
+        self._cudnn_benchmark = None
+
+    def __enter__(self):
+        self._cudnn_benchmark = torch.backends.cudnn.benchmark
+        torch.backends.cudnn.benchmark = self._chosen_cudnn_benchmark
+
+    def __exit__(self, exc_type, exc_value, exc_traceback):
+        torch.backends.cudnn.benchmark = self._cudnn_benchmark
+
+
+if torch.cuda.is_available():
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.allow_tf32 = True
+
+
+@contextlib.contextmanager
+def _inference_mode():
+    with InferenceContext(), torch.inference_mode(), torch.no_grad(), autocast():
+        yield
+
+
+def _clear_cuda_cache():
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+        torch.cuda.synchronize()
+
+
+def clean_models(model_key=None):
+    global models
+    model_keys = [model_key] if model_key is not None else models.keys()
+    for k in model_keys:
+        if k in models:
+            del models[k]
+    _clear_cuda_cache()
+    gc.collect()
+
+
+def _load_model(ckpt_path, device, use_small=False, model_type="text"):
+    if model_type == "text":
+        ConfigClass = GPTConfig
+        ModelClass = GPT
+    elif model_type == "coarse":
+        ConfigClass = GPTConfig
+        ModelClass = GPT
+    elif model_type == "fine":
+        ConfigClass = FineGPTConfig
+        ModelClass = FineGPT
+    else:
+        raise NotImplementedError()
+
+    # Force-remove Models to allow running on >12Gb GPU
+    # CF: Probably not needed anymore
+    #global models
+    #models.clear()
+    #gc.collect()
+    #torch.cuda.empty_cache()
+    # to here...
+
+    model_key = f"{model_type}_small" if use_small or USE_SMALL_MODELS else model_type
+    model_info = REMOTE_MODEL_PATHS[model_key]
+    if not os.path.exists(ckpt_path):
+        logger.info(f"{model_type} model not found, downloading into `{CACHE_DIR}`.")
+        ## added next two lines to make it super clear which model is being downloaded
+        remote_filename = hf_hub_url(model_info["repo_id"], model_info["file_name"])
+        print(f"Downloading {model_key} {model_info['repo_id']} remote model file {remote_filename} {model_info['file_name']} to {CACHE_DIR}")
+        _download(model_info["repo_id"], model_info["file_name"])
+        # add next line to make it super clear which model is being loaded
+    print(f"Loading {model_key} model from {ckpt_path} to {device}") # added
+    checkpoint = torch.load(ckpt_path, map_location=device)
+    # this is a hack
+    model_args = checkpoint["model_args"]
+    if "input_vocab_size" not in model_args:
+        model_args["input_vocab_size"] = model_args["vocab_size"]
+        model_args["output_vocab_size"] = model_args["vocab_size"]
+        del model_args["vocab_size"]
+    gptconf = ConfigClass(**checkpoint["model_args"])
+    model = ModelClass(gptconf)
+    state_dict = checkpoint["model"]
+    # fixup checkpoint
+    unwanted_prefix = "_orig_mod."
+    for k, v in list(state_dict.items()):
+        if k.startswith(unwanted_prefix):
+            state_dict[k[len(unwanted_prefix) :]] = state_dict.pop(k)
+    extra_keys = set(state_dict.keys()) - set(model.state_dict().keys())
+    extra_keys = set([k for k in extra_keys if not k.endswith(".attn.bias")])
+    missing_keys = set(model.state_dict().keys()) - set(state_dict.keys())
+    missing_keys = set([k for k in missing_keys if not k.endswith(".attn.bias")])
+    if len(extra_keys) != 0:
+        raise ValueError(f"extra keys found: {extra_keys}")
+    if len(missing_keys) != 0:
+        raise ValueError(f"missing keys: {missing_keys}")
+    model.load_state_dict(state_dict, strict=False)
+    n_params = model.get_num_params()
+    val_loss = checkpoint["best_val_loss"].item()
+    logger.info(f"model loaded: {round(n_params/1e6,1)}M params, {round(val_loss,3)} loss")
+    model.eval()
+    model.to(device)
+    del checkpoint, state_dict
+    _clear_cuda_cache()
+    if model_type == "text":
+        tokenizer = BertTokenizer.from_pretrained("bert-base-multilingual-cased")
+        return {
+            "model": model,
+            "tokenizer": tokenizer,
+        }
+    return model
+
+
+def _load_codec_model(device):
+    model = EncodecModel.encodec_model_24khz()
+    model.set_target_bandwidth(6.0)
+    model.eval()
+    model.to(device)
+    _clear_cuda_cache()
+    return model
+
+
+def load_model(use_gpu=True, use_small=False, force_reload=False, model_type="text"):
+    _load_model_f = funcy.partial(_load_model, model_type=model_type, use_small=use_small)
+    if model_type not in ("text", "coarse", "fine"):
+        raise NotImplementedError()
+    global models
+    global models_devices
+    device = grab_best_device(use_gpu=use_gpu)
+    model_key = f"{model_type}"
+    if OFFLOAD_CPU:
+        models_devices[model_key] = device
+        device = "cpu"
+    if model_key not in models or force_reload:
+        ckpt_path = _get_ckpt_path(model_type, use_small=use_small)
+        clean_models(model_key=model_key)
+        model = _load_model_f(ckpt_path, device)
+        models[model_key] = model
+    if model_type == "text":
+        models[model_key]["model"].to(device)
+    else:
+        models[model_key].to(device)
+    return models[model_key]
+
+
+def load_codec_model(use_gpu=True, force_reload=False):
+    global models
+    global models_devices
+    device = grab_best_device(use_gpu=use_gpu)
+    if device == "mps":
+        # encodec doesn't support mps
+        device = "cpu"
+    model_key = "codec"
+    if OFFLOAD_CPU:
+        models_devices[model_key] = device
+        device = "cpu"
+    if model_key not in models or force_reload:
+        clean_models(model_key=model_key)
+        model = _load_codec_model(device)
+        models[model_key] = model
+    models[model_key].to(device)
+    return models[model_key]
+
+
+def preload_models(
+    text_use_gpu=True,
+    text_use_small=False,
+    coarse_use_gpu=True,
+    coarse_use_small=False,
+    fine_use_gpu=True,
+    fine_use_small=False,
+    codec_use_gpu=True,
+    force_reload=False
+):
+    """Load all the necessary models for the pipeline."""
+    if grab_best_device() == "cpu" and (
+        text_use_gpu or coarse_use_gpu or fine_use_gpu or codec_use_gpu
+    ):
+        logger.warning("No GPU being used. Careful, inference might be very slow!")
+    _ = load_model(
+        model_type="text", use_gpu=text_use_gpu, use_small=text_use_small, force_reload=force_reload
+    )
+    _ = load_model(
+        model_type="coarse",
+        use_gpu=coarse_use_gpu,
+        use_small=coarse_use_small,
+        force_reload=force_reload,
+    )
+    _ = load_model(
+        model_type="fine", use_gpu=fine_use_gpu, use_small=fine_use_small, force_reload=force_reload
+    )
+    _ = load_codec_model(use_gpu=codec_use_gpu, force_reload=force_reload)
+
+
+####
+# Generation Functionality
+####
+
+
+def _tokenize(tokenizer, text):
+    return tokenizer.encode(text, add_special_tokens=False)
+
+
+def _detokenize(tokenizer, enc_text):
+    return tokenizer.decode(enc_text)
+
+
+def _normalize_whitespace(text):
+    return re.sub(r"\s+", " ", text).strip()
+
+
+TEXT_ENCODING_OFFSET = 10_048
+SEMANTIC_PAD_TOKEN = 10_000
+TEXT_PAD_TOKEN = 129_595
+SEMANTIC_INFER_TOKEN = 129_599
+
+
+def _load_history_prompt(history_prompt_input):
+    if isinstance(history_prompt_input, str) and history_prompt_input.endswith(".npz"):
+        history_prompt = np.load(history_prompt_input)
+    elif isinstance(history_prompt_input, str):
+        # make sure this works on non-ubuntu
+        history_prompt_input = os.path.join(*history_prompt_input.split("/"))
+#        if history_prompt_input not in ALLOWED_PROMPTS:
+#            raise ValueError("history prompt not found")
+        history_prompt = np.load(
+            os.path.join(CUR_PATH, "assets", "prompts", f"{history_prompt_input}.npz")
+        )
+    elif isinstance(history_prompt_input, dict):
+        assert("semantic_prompt" in history_prompt_input)
+        assert("coarse_prompt" in history_prompt_input)
+        assert("fine_prompt" in history_prompt_input)
+        history_prompt = history_prompt_input
+    else:
+        raise ValueError("history prompt format unrecognized")
+    return history_prompt
+
+
+def generate_text_semantic(
+    text,
+    history_prompt=None,
+    temp=0.7,
+    top_k=None,
+    top_p=None,
+    silent=False,
+    min_eos_p=0.2,
+    max_gen_duration_s=None,
+    allow_early_stop=True,
+    use_kv_caching=False,
+):
+    """Generate semantic tokens from text."""
+    assert isinstance(text, str)
+    text = _normalize_whitespace(text)
+    assert len(text.strip()) > 0
+    if history_prompt is not None:
+        history_prompt = _load_history_prompt(history_prompt)
+        semantic_history = history_prompt["semantic_prompt"]
+        assert (
+            isinstance(semantic_history, np.ndarray)
+            and len(semantic_history.shape) == 1
+            and len(semantic_history) > 0
+            and semantic_history.min() >= 0
+            and semantic_history.max() <= SEMANTIC_VOCAB_SIZE - 1
+        )
+    else:
+        semantic_history = None
+    # load models if not yet exist
+    global models
+    global models_devices
+    if "text" not in models:
+        preload_models()
+    model_container = models["text"]
+    model = model_container["model"]
+    tokenizer = model_container["tokenizer"]
+    encoded_text = np.array(_tokenize(tokenizer, text)) + TEXT_ENCODING_OFFSET
+    if OFFLOAD_CPU:
+        model.to(models_devices["text"])
+    device = next(model.parameters()).device
+    if len(encoded_text) > 256:
+        p = round((len(encoded_text) - 256) / len(encoded_text) * 100, 1)
+        logger.warning(f"warning, text too long, lopping of last {p}%")
+        encoded_text = encoded_text[:256]
+    encoded_text = np.pad(
+        encoded_text,
+        (0, 256 - len(encoded_text)),
+        constant_values=TEXT_PAD_TOKEN,
+        mode="constant",
+    )
+    if semantic_history is not None:
+        semantic_history = semantic_history.astype(np.int64)
+        # lop off if history is too long, pad if needed
+        semantic_history = semantic_history[-256:]
+        semantic_history = np.pad(
+            semantic_history,
+            (0, 256 - len(semantic_history)),
+            constant_values=SEMANTIC_PAD_TOKEN,
+            mode="constant",
+        )
+    else:
+        semantic_history = np.array([SEMANTIC_PAD_TOKEN] * 256)
+    x = torch.from_numpy(
+        np.hstack([
+            encoded_text, semantic_history, np.array([SEMANTIC_INFER_TOKEN])
+        ]).astype(np.int64)
+    )[None]
+    assert x.shape[1] == 256 + 256 + 1
+    with _inference_mode():
+        x = x.to(device)
+        n_tot_steps = 768
+        # custom tqdm updates since we don't know when eos will occur
+        pbar = tqdm.tqdm(disable=silent, total=100)
+        pbar_state = 0
+        tot_generated_duration_s = 0
+        kv_cache = None
+        for n in range(n_tot_steps):
+            if use_kv_caching and kv_cache is not None:
+                x_input = x[:, [-1]]
+            else:
+                x_input = x
+            logits, kv_cache = model(
+                x_input, merge_context=True, use_cache=use_kv_caching, past_kv=kv_cache
+            )
+            relevant_logits = logits[0, 0, :SEMANTIC_VOCAB_SIZE]
+            if allow_early_stop:
+                relevant_logits = torch.hstack(
+                    (relevant_logits, logits[0, 0, [SEMANTIC_PAD_TOKEN]])  # eos
+                )
+            if top_p is not None:
+                # faster to convert to numpy
+                original_device = relevant_logits.device
+                relevant_logits = relevant_logits.detach().cpu().type(torch.float32).numpy()
+                sorted_indices = np.argsort(relevant_logits)[::-1]
+                sorted_logits = relevant_logits[sorted_indices]
+                cumulative_probs = np.cumsum(softmax(sorted_logits))
+                sorted_indices_to_remove = cumulative_probs > top_p
+                sorted_indices_to_remove[1:] = sorted_indices_to_remove[:-1].copy()
+                sorted_indices_to_remove[0] = False
+                relevant_logits[sorted_indices[sorted_indices_to_remove]] = -np.inf
+                relevant_logits = torch.from_numpy(relevant_logits)
+                relevant_logits = relevant_logits.to(original_device)
+            if top_k is not None:
+                v, _ = torch.topk(relevant_logits, min(top_k, relevant_logits.size(-1)))
+                relevant_logits[relevant_logits < v[-1]] = -float("Inf")
+            probs = F.softmax(relevant_logits / temp, dim=-1)
+            # multinomial bugged on mps: shuttle to cpu if necessary
+            inf_device = probs.device
+            if probs.device.type == "mps":
+                probs = probs.to("cpu")
+            item_next = torch.multinomial(probs, num_samples=1)
+            probs = probs.to(inf_device)
+            item_next = item_next.to(inf_device)
+            if allow_early_stop and (
+                item_next == SEMANTIC_VOCAB_SIZE
+                or (min_eos_p is not None and probs[-1] >= min_eos_p)
+            ):
+                # eos found, so break
+                pbar.update(100 - pbar_state)
+                break
+            x = torch.cat((x, item_next[None]), dim=1)
+            tot_generated_duration_s += 1 / SEMANTIC_RATE_HZ
+            if max_gen_duration_s is not None and tot_generated_duration_s > max_gen_duration_s:
+                pbar.update(100 - pbar_state)
+                break
+            if n == n_tot_steps - 1:
+                pbar.update(100 - pbar_state)
+                break
+            del logits, relevant_logits, probs, item_next
+            req_pbar_state = np.min([100, int(round(100 * n / n_tot_steps))])
+            if req_pbar_state > pbar_state:
+                pbar.update(req_pbar_state - pbar_state)
+            pbar_state = req_pbar_state
+        pbar.close()
+        out = x.detach().cpu().numpy().squeeze()[256 + 256 + 1 :]
+    if OFFLOAD_CPU:
+        model.to("cpu")
+    assert all(0 <= out) and all(out < SEMANTIC_VOCAB_SIZE)
+    _clear_cuda_cache()
+    return out
+
+
+def _flatten_codebooks(arr, offset_size=CODEBOOK_SIZE):
+    assert len(arr.shape) == 2
+    arr = arr.copy()
+    if offset_size is not None:
+        for n in range(1, arr.shape[0]):
+            arr[n, :] += offset_size * n
+    flat_arr = arr.ravel("F")
+    return flat_arr
+
+
+COARSE_SEMANTIC_PAD_TOKEN = 12_048
+COARSE_INFER_TOKEN = 12_050
+
+
+def generate_coarse(
+    x_semantic,
+    history_prompt=None,
+    temp=0.7,
+    top_k=None,
+    top_p=None,
+    silent=False,
+    max_coarse_history=630,  # min 60 (faster), max 630 (more context)
+    sliding_window_len=60,
+    use_kv_caching=False,
+):
+    """Generate coarse audio codes from semantic tokens."""
+# CF: Uncommented because it breaks swap voice more than once
+#    assert (
+#        isinstance(x_semantic, np.ndarray)
+#        and len(x_semantic.shape) == 1
+#        and len(x_semantic) > 0
+#        and x_semantic.min() >= 0
+#        and x_semantic.max() <= SEMANTIC_VOCAB_SIZE - 1
+#    )
+    assert 60 <= max_coarse_history <= 630
+    assert max_coarse_history + sliding_window_len <= 1024 - 256
+    semantic_to_coarse_ratio = COARSE_RATE_HZ / SEMANTIC_RATE_HZ * N_COARSE_CODEBOOKS
+    max_semantic_history = int(np.floor(max_coarse_history / semantic_to_coarse_ratio))
+    if history_prompt is not None:
+        history_prompt = _load_history_prompt(history_prompt)
+        x_semantic_history = history_prompt["semantic_prompt"]
+        x_coarse_history = history_prompt["coarse_prompt"]
+        assert (
+            isinstance(x_semantic_history, np.ndarray)
+            and len(x_semantic_history.shape) == 1
+            and len(x_semantic_history) > 0
+            and x_semantic_history.min() >= 0
+            and x_semantic_history.max() <= SEMANTIC_VOCAB_SIZE - 1
+            and isinstance(x_coarse_history, np.ndarray)
+            and len(x_coarse_history.shape) == 2
+            and x_coarse_history.shape[0] == N_COARSE_CODEBOOKS
+            and x_coarse_history.shape[-1] >= 0
+            and x_coarse_history.min() >= 0
+            and x_coarse_history.max() <= CODEBOOK_SIZE - 1
+            #and (
+            #    round(x_coarse_history.shape[-1] / len(x_semantic_history), 1)
+            #    == round(semantic_to_coarse_ratio / N_COARSE_CODEBOOKS, 1)
+            #)
+        )
+        x_coarse_history = _flatten_codebooks(x_coarse_history) + SEMANTIC_VOCAB_SIZE
+        # trim histories correctly
+        n_semantic_hist_provided = np.min(
+            [
+                max_semantic_history,
+                len(x_semantic_history) - len(x_semantic_history) % 2,
+                int(np.floor(len(x_coarse_history) / semantic_to_coarse_ratio)),
+            ]
+        )
+        n_coarse_hist_provided = int(round(n_semantic_hist_provided * semantic_to_coarse_ratio))
+        x_semantic_history = x_semantic_history[-n_semantic_hist_provided:].astype(np.int32)
+        x_coarse_history = x_coarse_history[-n_coarse_hist_provided:].astype(np.int32)
+        # TODO: bit of a hack for time alignment (sounds better)
+        x_coarse_history = x_coarse_history[:-2]
+    else:
+        x_semantic_history = np.array([], dtype=np.int32)
+        x_coarse_history = np.array([], dtype=np.int32)
+    # load models if not yet exist
+    global models
+    global models_devices
+    if "coarse" not in models:
+        preload_models()
+    model = models["coarse"]
+    if OFFLOAD_CPU:
+        model.to(models_devices["coarse"])
+    device = next(model.parameters()).device
+    # start loop
+    n_steps = int(
+        round(
+            np.floor(len(x_semantic) * semantic_to_coarse_ratio / N_COARSE_CODEBOOKS)
+            * N_COARSE_CODEBOOKS
+        )
+    )
+    assert n_steps > 0 and n_steps % N_COARSE_CODEBOOKS == 0
+    x_semantic = np.hstack([x_semantic_history, x_semantic]).astype(np.int32)
+    x_coarse = x_coarse_history.astype(np.int32)
+    base_semantic_idx = len(x_semantic_history)
+    with _inference_mode():
+        x_semantic_in = torch.from_numpy(x_semantic)[None].to(device)
+        x_coarse_in = torch.from_numpy(x_coarse)[None].to(device)
+        n_window_steps = int(np.ceil(n_steps / sliding_window_len))
+        n_step = 0
+        for _ in tqdm.tqdm(range(n_window_steps), total=n_window_steps, disable=silent):
+            semantic_idx = base_semantic_idx + int(round(n_step / semantic_to_coarse_ratio))
+            # pad from right side
+            x_in = x_semantic_in[:, np.max([0, semantic_idx - max_semantic_history]) :]
+            x_in = x_in[:, :256]
+            x_in = F.pad(
+                x_in,
+                (0, 256 - x_in.shape[-1]),
+                "constant",
+                COARSE_SEMANTIC_PAD_TOKEN,
+            )
+            x_in = torch.hstack(
+                [
+                    x_in,
+                    torch.tensor([COARSE_INFER_TOKEN])[None].to(device),
+                    x_coarse_in[:, -max_coarse_history:],
+                ]
+            )
+            kv_cache = None
+            for _ in range(sliding_window_len):
+                if n_step >= n_steps:
+                    continue
+                is_major_step = n_step % N_COARSE_CODEBOOKS == 0
+
+                if use_kv_caching and kv_cache is not None:
+                    x_input = x_in[:, [-1]]
+                else:
+                    x_input = x_in
+
+                logits, kv_cache = model(x_input, use_cache=use_kv_caching, past_kv=kv_cache)
+                logit_start_idx = (
+                    SEMANTIC_VOCAB_SIZE + (1 - int(is_major_step)) * CODEBOOK_SIZE
+                )
+                logit_end_idx = (
+                    SEMANTIC_VOCAB_SIZE + (2 - int(is_major_step)) * CODEBOOK_SIZE
+                )
+                relevant_logits = logits[0, 0, logit_start_idx:logit_end_idx]
+                if top_p is not None:
+                    # faster to convert to numpy
+                    original_device = relevant_logits.device
+                    relevant_logits = relevant_logits.detach().cpu().type(torch.float32).numpy()
+                    sorted_indices = np.argsort(relevant_logits)[::-1]
+                    sorted_logits = relevant_logits[sorted_indices]
+                    cumulative_probs = np.cumsum(softmax(sorted_logits))
+                    sorted_indices_to_remove = cumulative_probs > top_p
+                    sorted_indices_to_remove[1:] = sorted_indices_to_remove[:-1].copy()
+                    sorted_indices_to_remove[0] = False
+                    relevant_logits[sorted_indices[sorted_indices_to_remove]] = -np.inf
+                    relevant_logits = torch.from_numpy(relevant_logits)
+                    relevant_logits = relevant_logits.to(original_device)
+                if top_k is not None:
+                    v, _ = torch.topk(relevant_logits, min(top_k, relevant_logits.size(-1)))
+                    relevant_logits[relevant_logits < v[-1]] = -float("Inf")
+                probs = F.softmax(relevant_logits / temp, dim=-1)
+                # multinomial bugged on mps: shuttle to cpu if necessary
+                inf_device = probs.device
+                if probs.device.type == "mps":
+                    probs = probs.to("cpu")
+                item_next = torch.multinomial(probs, num_samples=1)
+                probs = probs.to(inf_device)
+                item_next = item_next.to(inf_device)
+                item_next += logit_start_idx
+                x_coarse_in = torch.cat((x_coarse_in, item_next[None]), dim=1)
+                x_in = torch.cat((x_in, item_next[None]), dim=1)
+                del logits, relevant_logits, probs, item_next
+                n_step += 1
+            del x_in
+        del x_semantic_in
+    if OFFLOAD_CPU:
+        model.to("cpu")
+    gen_coarse_arr = x_coarse_in.detach().cpu().numpy().squeeze()[len(x_coarse_history) :]
+    del x_coarse_in
+    assert len(gen_coarse_arr) == n_steps
+    gen_coarse_audio_arr = gen_coarse_arr.reshape(-1, N_COARSE_CODEBOOKS).T - SEMANTIC_VOCAB_SIZE
+    for n in range(1, N_COARSE_CODEBOOKS):
+        gen_coarse_audio_arr[n, :] -= n * CODEBOOK_SIZE
+    _clear_cuda_cache()
+    return gen_coarse_audio_arr
+
+
+def generate_fine(
+    x_coarse_gen,
+    history_prompt=None,
+    temp=0.5,
+    silent=True,
+):
+    """Generate full audio codes from coarse audio codes."""
+    assert (
+        isinstance(x_coarse_gen, np.ndarray)
+        and len(x_coarse_gen.shape) == 2
+        and 1 <= x_coarse_gen.shape[0] <= N_FINE_CODEBOOKS - 1
+        and x_coarse_gen.shape[1] > 0
+        and x_coarse_gen.min() >= 0
+        and x_coarse_gen.max() <= CODEBOOK_SIZE - 1
+    )
+    if history_prompt is not None:
+        history_prompt = _load_history_prompt(history_prompt)
+        x_fine_history = history_prompt["fine_prompt"]
+        assert (
+            isinstance(x_fine_history, np.ndarray)
+            and len(x_fine_history.shape) == 2
+            and x_fine_history.shape[0] == N_FINE_CODEBOOKS
+            and x_fine_history.shape[1] >= 0
+            and x_fine_history.min() >= 0
+            and x_fine_history.max() <= CODEBOOK_SIZE - 1
+        )
+    else:
+        x_fine_history = None
+    n_coarse = x_coarse_gen.shape[0]
+    # load models if not yet exist
+    global models
+    global models_devices
+    if "fine" not in models:
+        preload_models()
+    model = models["fine"]
+    if OFFLOAD_CPU:
+        model.to(models_devices["fine"])
+    device = next(model.parameters()).device
+    # make input arr
+    in_arr = np.vstack(
+        [
+            x_coarse_gen,
+            np.zeros((N_FINE_CODEBOOKS - n_coarse, x_coarse_gen.shape[1]))
+            + CODEBOOK_SIZE,  # padding
+        ]
+    ).astype(np.int32)
+    # prepend history if available (max 512)
+    if x_fine_history is not None:
+        x_fine_history = x_fine_history.astype(np.int32)
+        in_arr = np.hstack(
+            [
+                x_fine_history[:, -512:].astype(np.int32),
+                in_arr,
+            ]
+        )
+        n_history = x_fine_history[:, -512:].shape[1]
+    else:
+        n_history = 0
+    n_remove_from_end = 0
+    # need to pad if too short (since non-causal model)
+    if in_arr.shape[1] < 1024:
+        n_remove_from_end = 1024 - in_arr.shape[1]
+        in_arr = np.hstack(
+            [
+                in_arr,
+                np.zeros((N_FINE_CODEBOOKS, n_remove_from_end), dtype=np.int32) + CODEBOOK_SIZE,
+            ]
+        )
+    # we can be lazy about fractional loop and just keep overwriting codebooks
+    n_loops = np.max([0, int(np.ceil((x_coarse_gen.shape[1] - (1024 - n_history)) / 512))]) + 1
+    with _inference_mode():
+        in_arr = torch.tensor(in_arr.T).to(device)
+        for n in tqdm.tqdm(range(n_loops), disable=silent):
+            start_idx = np.min([n * 512, in_arr.shape[0] - 1024])
+            start_fill_idx = np.min([n_history + n * 512, in_arr.shape[0] - 512])
+            rel_start_fill_idx = start_fill_idx - start_idx
+            in_buffer = in_arr[start_idx : start_idx + 1024, :][None]
+            for nn in range(n_coarse, N_FINE_CODEBOOKS):
+                logits = model(nn, in_buffer)
+                if temp is None:
+                    relevant_logits = logits[0, rel_start_fill_idx:, :CODEBOOK_SIZE]
+                    codebook_preds = torch.argmax(relevant_logits, -1)
+                else:
+                    relevant_logits = logits[0, :, :CODEBOOK_SIZE] / temp
+                    probs = F.softmax(relevant_logits, dim=-1)
+                    # multinomial bugged on mps: shuttle to cpu if necessary
+                    inf_device = probs.device
+                    if probs.device.type == "mps":
+                        probs = probs.to("cpu")
+                    codebook_preds = torch.hstack(
+                        [
+                            torch.multinomial(probs[nnn], num_samples=1).to(inf_device)
+                            for nnn in range(rel_start_fill_idx, 1024)
+                        ]
+                    )
+                in_buffer[0, rel_start_fill_idx:, nn] = codebook_preds
+                del logits, codebook_preds
+            # transfer over info into model_in and convert to numpy
+            for nn in range(n_coarse, N_FINE_CODEBOOKS):
+                in_arr[
+                    start_fill_idx : start_fill_idx + (1024 - rel_start_fill_idx), nn
+                ] = in_buffer[0, rel_start_fill_idx:, nn]
+            del in_buffer
+        gen_fine_arr = in_arr.detach().cpu().numpy().squeeze().T
+        del in_arr
+    if OFFLOAD_CPU:
+        model.to("cpu")
+    gen_fine_arr = gen_fine_arr[:, n_history:]
+    if n_remove_from_end > 0:
+        gen_fine_arr = gen_fine_arr[:, :-n_remove_from_end]
+    assert gen_fine_arr.shape[-1] == x_coarse_gen.shape[-1]
+    _clear_cuda_cache()
+    return gen_fine_arr
+
+
+def codec_decode(fine_tokens):
+    """Turn quantized audio codes into audio array using encodec."""
+    # load models if not yet exist
+    global models
+    global models_devices
+    if "codec" not in models:
+        preload_models()
+    model = models["codec"]
+    if OFFLOAD_CPU:
+        model.to(models_devices["codec"])
+    device = next(model.parameters()).device
+    arr = torch.from_numpy(fine_tokens)[None]
+    arr = arr.to(device)
+    arr = arr.transpose(0, 1)
+    emb = model.quantizer.decode(arr)
+    out = model.decoder(emb)
+    audio_arr = out.detach().cpu().numpy().squeeze()
+    del arr, emb, out
+    if OFFLOAD_CPU:
+        model.to("cpu")
+    return audio_arr

bark/hubert/customtokenizer.py

@@ -0,0 +1,195 @@
+"""
+Custom tokenizer model.
+Author: https://www.github.com/gitmylo/
+License: MIT
+"""
+
+import json
+import os.path
+from zipfile import ZipFile
+
+import numpy
+import torch
+from torch import nn, optim
+from torch.serialization import MAP_LOCATION
+from tqdm.auto import tqdm
+
+
+class CustomTokenizer(nn.Module):
+    def __init__(self, hidden_size=1024, input_size=768, output_size=10000, version=0):
+        super(CustomTokenizer, self).__init__()
+        next_size = input_size
+        if version == 0:
+            self.lstm = nn.LSTM(input_size, hidden_size, 2, batch_first=True)
+            next_size = hidden_size
+        if version == 1:
+            self.lstm = nn.LSTM(input_size, hidden_size, 2, batch_first=True)
+            self.intermediate = nn.Linear(hidden_size, 4096)
+            next_size = 4096
+
+        self.fc = nn.Linear(next_size, output_size)
+        self.softmax = nn.LogSoftmax(dim=1)
+        self.optimizer: optim.Optimizer = None
+        self.lossfunc = nn.CrossEntropyLoss()
+        self.input_size = input_size
+        self.hidden_size = hidden_size
+        self.output_size = output_size
+        self.version = version
+
+    def forward(self, x):
+        x, _ = self.lstm(x)
+        if self.version == 1:
+            x = self.intermediate(x)
+        x = self.fc(x)
+        x = self.softmax(x)
+        return x
+
+    @torch.no_grad()
+    def get_token(self, x):
+        """
+        Used to get the token for the first
+        :param x: An array with shape (N, input_size) where N is a whole number greater or equal to 1, and input_size is the input size used when creating the model.
+        :return: An array with shape (N,) where N is the same as N from the input. Every number in the array is a whole number in range 0...output_size - 1 where output_size is the output size used when creating the model.
+        """
+        return torch.argmax(self(x), dim=1)
+
+    def prepare_training(self):
+        self.optimizer = optim.Adam(self.parameters(), 0.001)
+
+    def train_step(self, x_train, y_train, log_loss=False):
+        # y_train = y_train[:-1]
+        # y_train = y_train[1:]
+
+        optimizer = self.optimizer
+        lossfunc = self.lossfunc
+        # Zero the gradients
+        self.zero_grad()
+
+        # Forward pass
+        y_pred = self(x_train)
+
+        y_train_len = len(y_train)
+        y_pred_len = y_pred.shape[0]
+
+        if y_train_len > y_pred_len:
+            diff = y_train_len - y_pred_len
+            y_train = y_train[diff:]
+        elif y_train_len < y_pred_len:
+            diff = y_pred_len - y_train_len
+            y_pred = y_pred[:-diff, :]
+
+        y_train_hot = torch.zeros(len(y_train), self.output_size)
+        y_train_hot[range(len(y_train)), y_train] = 1
+        y_train_hot = y_train_hot.to('cuda')
+
+        # Calculate the loss
+        loss = lossfunc(y_pred, y_train_hot)
+
+        # Print loss
+        if log_loss:
+            print('Loss', loss.item())
+
+        # Backward pass
+        loss.backward()
+
+        # Update the weights
+        optimizer.step()
+
+    def save(self, path):
+        info_path = '.'.join(os.path.basename(path).split('.')[:-1]) + '/.info'
+        torch.save(self.state_dict(), path)
+        data_from_model = Data(self.input_size, self.hidden_size, self.output_size, self.version)
+        with ZipFile(path, 'a') as model_zip:
+            model_zip.writestr(info_path, data_from_model.save())
+            model_zip.close()
+
+    @staticmethod
+    def load_from_checkpoint(path, map_location: MAP_LOCATION = None):
+        old = True
+        with ZipFile(path) as model_zip:
+            filesMatch = [file for file in model_zip.namelist() if file.endswith('/.info')]
+            file = filesMatch[0] if filesMatch else None
+            if file:
+                old = False
+                print(f"Loading Custom Hubert Tokenizer {path}")
+                data_from_model = Data.load(model_zip.read(file).decode('utf-8'))
+            model_zip.close()
+        if old:
+            model = CustomTokenizer()
+        else:
+            model = CustomTokenizer(data_from_model.hidden_size, data_from_model.input_size, data_from_model.output_size, data_from_model.version)
+        model.load_state_dict(torch.load(path))
+        if map_location:
+            model = model.to(map_location)
+        return model
+
+
+
+class Data:
+    input_size: int
+    hidden_size: int
+    output_size: int
+    version: int
+
+    def __init__(self, input_size=768, hidden_size=1024, output_size=10000, version=0):
+        self.input_size = input_size
+        self.hidden_size = hidden_size
+        self.output_size = output_size
+        self.version = version
+
+    @staticmethod
+    def load(string):
+        data = json.loads(string)
+        return Data(data['input_size'], data['hidden_size'], data['output_size'], data['version'])
+
+    def save(self):
+        data = {
+            'input_size': self.input_size,
+            'hidden_size': self.hidden_size,
+            'output_size': self.output_size,
+            'version': self.version,
+        }
+        return json.dumps(data)
+
+
+def auto_train(data_path, save_path='model.pth', load_model: str | None = None, save_epochs=1, max_epochs=14):
+    data_x, data_y = [], []
+
+    if load_model and os.path.isfile(load_model):
+        print('Loading model from', load_model)
+        model_training = CustomTokenizer.load_from_checkpoint(load_model, 'cuda')
+    else:
+        print('Creating new model.')
+        model_training = CustomTokenizer(version=1).to('cuda')  # Settings for the model to run without lstm
+    save_path = os.path.join(data_path, save_path)
+    base_save_path = '.'.join(save_path.split('.')[:-1])
+
+    sem_string = '_semantic.npy'
+    feat_string = '_semantic_features.npy'
+
+    ready = os.path.join(data_path, 'ready')
+    for input_file in os.listdir(ready):
+        full_path = os.path.join(ready, input_file)
+        if input_file.endswith(sem_string):
+            data_y.append(numpy.load(full_path))
+        elif input_file.endswith(feat_string):
+            data_x.append(numpy.load(full_path))
+    model_training.prepare_training()
+
+    epoch = 1
+    with tqdm(total=((len(data_x) * len(data_y)) / 50) * save_epochs) as pbar1:
+        while epoch <= max_epochs:
+            for i in range(save_epochs):
+                j = 0
+                for x, y in zip(data_x, data_y):
+                    model_training.train_step(torch.tensor(x).to('cuda'), torch.tensor(y).to('cuda'), j % 50 == 0)  # Print loss every 50 steps
+                    j += 1
+                    pbar1.update()
+
+            save_p = save_path
+            save_p_2 = f'{base_save_path}_epoch_{epoch}.pth'
+            model_training.save(save_p)
+            model_training.save(save_p_2)
+            print(f'Epoch {epoch} completed')
+            epoch += 1
+        print(f'Done training for {max_epochs} epochs!')

bark/hubert/hubert_manager.py

@@ -0,0 +1,48 @@
+import os.path
+import shutil
+import urllib.request
+
+import huggingface_hub
+
+
+class HuBERTManager:
+
+
+    @staticmethod
+    def make_sure_hubert_installed(download_url: str = 'https://dl.fbaipublicfiles.com/hubert/hubert_base_ls960.pt', file_name: str = 'hubert.pt'):
+        install_dir = os.path.join('models', 'hubert')
+        if not os.path.isdir(install_dir):
+            os.makedirs(install_dir, exist_ok=True)
+        install_file = os.path.join(install_dir, file_name)
+        if not os.path.isfile(install_file):
+            print(f'Downloading HuBERT base model from {download_url}')
+            urllib.request.urlretrieve(download_url, install_file)
+            print('Downloaded HuBERT')
+        return install_file
+
+
+    @staticmethod
+    def make_sure_tokenizer_installed(model: str = 'quantifier_hubert_base_ls960_14.pth', repo: str = 'GitMylo/bark-voice-cloning', tokenizer_lang: str = 'en'):
+        local_file = tokenizer_lang + '_tokenizer.pth'
+        install_dir = os.path.join('models', 'hubert')
+        if not os.path.isdir(install_dir):
+            os.makedirs(install_dir, exist_ok=True)
+        install_file = os.path.join(install_dir, local_file)
+        if not os.path.isfile(install_file):
+            # refactor to use lists
+            if tokenizer_lang == 'en':
+                repo = 'GitMylo/bark-voice-cloning'
+                model = 'quantifier_hubert_base_ls960_14.pth'
+            elif tokenizer_lang == 'de':
+                repo = 'CountFloyd/bark-voice-cloning-german-HuBERT-quantizer'
+                model = 'german-HuBERT-quantizer_14_epoch.pth'
+            elif tokenizer_lang == 'pl':
+                repo = 'Hobis/bark-voice-cloning-polish-HuBERT-quantizer'
+                model = 'polish-HuBERT-quantizer_8_epoch.pth'
+            else:
+                raise 'Unknown Tokenizer Language!'
+            print(f'{local_file} not found. Downloading HuBERT custom tokenizer')
+            huggingface_hub.hf_hub_download(repo, model, local_dir=install_dir, local_dir_use_symlinks=False)
+            shutil.move(os.path.join(install_dir, model), install_file)
+            print('Downloaded tokenizer')
+        return install_file

bark/hubert/pre_kmeans_hubert.py

@@ -0,0 +1,107 @@
+"""
+Modified HuBERT model without kmeans.
+Original author: https://github.com/lucidrains/
+Modified by: https://www.github.com/gitmylo/
+License: MIT
+"""
+
+# Modified code from https://github.com/lucidrains/audiolm-pytorch/blob/main/audiolm_pytorch/hubert_kmeans.py
+
+from pathlib import Path
+
+import torch
+from torch import nn
+from einops import pack, unpack
+
+import fairseq
+
+from torchaudio.functional import resample
+
+from audiolm_pytorch.utils import curtail_to_multiple
+
+import logging
+logging.root.setLevel(logging.ERROR)
+
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    return val if exists(val) else d
+
+
+class CustomHubert(nn.Module):
+    """
+    checkpoint and kmeans can be downloaded at https://github.com/facebookresearch/fairseq/tree/main/examples/hubert
+    or you can train your own
+    """
+
+    def __init__(
+        self,
+        checkpoint_path,
+        target_sample_hz=16000,
+        seq_len_multiple_of=None,
+        output_layer=9,
+        device=None
+    ):
+        super().__init__()
+        self.target_sample_hz = target_sample_hz
+        self.seq_len_multiple_of = seq_len_multiple_of
+        self.output_layer = output_layer
+
+        if device is not None:
+            self.to(device)
+
+        model_path = Path(checkpoint_path)
+
+        assert model_path.exists(), f'path {checkpoint_path} does not exist'
+
+        print(f"Loading Hubert {checkpoint_path}")
+        checkpoint = torch.load(checkpoint_path)
+        load_model_input = {checkpoint_path: checkpoint}
+        model, *_ = fairseq.checkpoint_utils.load_model_ensemble_and_task(load_model_input)
+
+        if device is not None:
+            model[0].to(device)
+
+        self.model = model[0]
+        self.model.eval()
+
+    @property
+    def groups(self):
+        return 1
+
+    @torch.no_grad()
+    def forward(
+        self,
+        wav_input,
+        flatten=True,
+        input_sample_hz=None
+    ):
+        device = wav_input.device
+
+        if exists(input_sample_hz):
+            wav_input = resample(wav_input, input_sample_hz, self.target_sample_hz)
+
+        if exists(self.seq_len_multiple_of):
+            wav_input = curtail_to_multiple(wav_input, self.seq_len_multiple_of)
+
+        embed = self.model(
+            wav_input,
+            features_only=True,
+            mask=False,  # thanks to @maitycyrus for noticing that mask is defaulted to True in the fairseq code
+            output_layer=self.output_layer
+        )
+
+        embed, packed_shape = pack([embed['x']], '* d')
+
+        # codebook_indices = self.kmeans.predict(embed.cpu().detach().numpy())
+
+        codebook_indices = torch.from_numpy(embed.cpu().detach().numpy()).to(device)  # .long()
+
+        if flatten:
+            return codebook_indices
+
+        codebook_indices, = unpack(codebook_indices, packed_shape, '*')
+        return codebook_indices

bark/model.py

@@ -0,0 +1,218 @@
+"""
+Much of this code is adapted from Andrej Karpathy's NanoGPT
+(https://github.com/karpathy/nanoGPT)
+"""
+import math
+from dataclasses import dataclass
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+class LayerNorm(nn.Module):
+    """ LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False """
+
+    def __init__(self, ndim, bias):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(ndim))
+        self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None
+
+    def forward(self, input):
+        return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)
+
+class CausalSelfAttention(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        # key, query, value projections for all heads, but in a batch
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
+        # output projection
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
+        # regularization
+        self.attn_dropout = nn.Dropout(config.dropout)
+        self.resid_dropout = nn.Dropout(config.dropout)
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.dropout = config.dropout
+        # flash attention make GPU go brrrrr but support is only in PyTorch nightly and still a bit scary
+        self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
+        if not self.flash:
+            # print("WARNING: using slow attention. Flash Attention atm needs PyTorch nightly and dropout=0.0")
+            # causal mask to ensure that attention is only applied to the left in the input sequence
+            self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size))
+                                        .view(1, 1, config.block_size, config.block_size))
+
+    def forward(self, x, past_kv=None, use_cache=False):
+        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
+
+        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
+        q, k ,v  = self.c_attn(x).split(self.n_embd, dim=2)
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+
+        if past_kv is not None:
+            past_key = past_kv[0]
+            past_value = past_kv[1]
+            k = torch.cat((past_key, k), dim=-2)
+            v = torch.cat((past_value, v), dim=-2)
+
+        FULL_T = k.shape[-2]
+
+        if use_cache is True:
+            present = (k, v)
+        else:
+            present = None
+
+        # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
+        if self.flash:
+            # efficient attention using Flash Attention CUDA kernels
+            if past_kv is not None:
+                # When `past_kv` is provided, we're doing incremental decoding and `q.shape[2] == 1`: q only contains
+                # the query for the last token. scaled_dot_product_attention interprets this as the first token in the
+                # sequence, so if is_causal=True it will mask out all attention from it. This is not what we want, so 
+                # to work around this we set is_causal=False.
+                is_causal = False
+            else:
+                is_causal = True
+
+            y = torch.nn.functional.scaled_dot_product_attention(q, k, v, dropout_p=self.dropout, is_causal=is_causal)
+        else:
+            # manual implementation of attention
+            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+            att = att.masked_fill(self.bias[:,:,FULL_T-T:FULL_T,:FULL_T] == 0, float('-inf'))
+            att = F.softmax(att, dim=-1)
+            att = self.attn_dropout(att)
+            y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
+
+        # output projection
+        y = self.resid_dropout(self.c_proj(y))
+        return (y, present)
+
+class MLP(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.c_fc    = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)
+        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)
+        self.dropout = nn.Dropout(config.dropout)
+        self.gelu = nn.GELU()
+
+    def forward(self, x):
+        x = self.c_fc(x)
+        x = self.gelu(x)
+        x = self.c_proj(x)
+        x = self.dropout(x)
+        return x
+
+class Block(nn.Module):
+
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.ln_1 = LayerNorm(config.n_embd, bias=config.bias)
+        self.attn = CausalSelfAttention(config)
+        self.ln_2 = LayerNorm(config.n_embd, bias=config.bias)
+        self.mlp = MLP(config)
+        self.layer_idx = layer_idx
+
+    def forward(self, x, past_kv=None, use_cache=False):
+        attn_output, prev_kvs = self.attn(self.ln_1(x), past_kv=past_kv, use_cache=use_cache)
+        x = x + attn_output
+        x = x + self.mlp(self.ln_2(x))
+        return (x, prev_kvs)
+
+@dataclass
+class GPTConfig:
+    block_size: int = 1024
+    input_vocab_size: int = 10_048
+    output_vocab_size: int = 10_048
+    n_layer: int = 12
+    n_head: int = 12
+    n_embd: int = 768
+    dropout: float = 0.0
+    bias: bool = True # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster
+
+class GPT(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.input_vocab_size is not None
+        assert config.output_vocab_size is not None
+        assert config.block_size is not None
+        self.config = config
+
+        self.transformer = nn.ModuleDict(dict(
+            wte = nn.Embedding(config.input_vocab_size, config.n_embd),
+            wpe = nn.Embedding(config.block_size, config.n_embd),
+            drop = nn.Dropout(config.dropout),
+            h = nn.ModuleList([Block(config, idx) for idx in range(config.n_layer)]),
+            ln_f = LayerNorm(config.n_embd, bias=config.bias),
+        ))
+        self.lm_head = nn.Linear(config.n_embd, config.output_vocab_size, bias=False)
+
+    def get_num_params(self, non_embedding=True):
+        """
+        Return the number of parameters in the model.
+        For non-embedding count (default), the position embeddings get subtracted.
+        The token embeddings would too, except due to the parameter sharing these
+        params are actually used as weights in the final layer, so we include them.
+        """
+        n_params = sum(p.numel() for p in self.parameters())
+        if non_embedding:
+            n_params -= self.transformer.wte.weight.numel()
+            n_params -= self.transformer.wpe.weight.numel()
+        return n_params
+
+    def forward(self, idx, merge_context=False, past_kv=None, position_ids=None, use_cache=False):
+        device = idx.device
+        b, t = idx.size()
+        if past_kv is not None:
+            assert t == 1
+            tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
+        else:
+            if merge_context:
+                assert(idx.shape[1] >= 256+256+1)
+                t = idx.shape[1] - 256
+            else:
+                assert t <= self.config.block_size, f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"
+
+            # forward the GPT model itself
+            if merge_context:
+                tok_emb = torch.cat([
+                    self.transformer.wte(idx[:,:256]) + self.transformer.wte(idx[:,256:256+256]),
+                    self.transformer.wte(idx[:,256+256:])
+                ], dim=1)
+            else:
+                tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
+
+        if past_kv is None:
+            past_length = 0
+            past_kv = tuple([None] * len(self.transformer.h))
+        else:
+            past_length = past_kv[0][0].size(-2)
+
+        if position_ids is None:
+            position_ids = torch.arange(past_length, t + past_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0) # shape (1, t)
+            assert position_ids.shape == (1, t)
+
+        pos_emb = self.transformer.wpe(position_ids) # position embeddings of shape (1, t, n_embd)
+
+        x = self.transformer.drop(tok_emb + pos_emb)
+
+        new_kv = () if use_cache else None
+
+        for i, (block, past_layer_kv) in enumerate(zip(self.transformer.h, past_kv)):
+            x, kv = block(x, past_kv=past_layer_kv, use_cache=use_cache)
+
+            if use_cache:
+                new_kv = new_kv + (kv,)
+
+        x = self.transformer.ln_f(x)
+
+        # inference-time mini-optimization: only forward the lm_head on the very last position
+        logits = self.lm_head(x[:, [-1], :]) # note: using list [-1] to preserve the time dim
+
+        return (logits, new_kv)

bark/model_fine.py

@@ -0,0 +1,149 @@
+"""
+Much of this code is adapted from Andrej Karpathy's NanoGPT
+(https://github.com/karpathy/nanoGPT)
+"""
+from dataclasses import dataclass
+import math
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+from .model import GPT, GPTConfig, MLP
+
+
+class NonCausalSelfAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        # key, query, value projections for all heads, but in a batch
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
+        # output projection
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
+        # regularization
+        self.attn_dropout = nn.Dropout(config.dropout)
+        self.resid_dropout = nn.Dropout(config.dropout)
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.dropout = config.dropout
+        # flash attention make GPU go brrrrr but support is only in PyTorch nightly and still a bit scary
+        self.flash = (
+            hasattr(torch.nn.functional, "scaled_dot_product_attention") and self.dropout == 0.0
+        )
+
+    def forward(self, x):
+        B, T, C = x.size()  # batch size, sequence length, embedding dimensionality (n_embd)
+
+        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
+        q, k, v = self.c_attn(x).split(self.n_embd, dim=2)
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)  # (B, nh, T, hs)
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)  # (B, nh, T, hs)
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)  # (B, nh, T, hs)
+
+        # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
+        if self.flash:
+            # efficient attention using Flash Attention CUDA kernels
+            y = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v, attn_mask=None, dropout_p=self.dropout, is_causal=False
+            )
+        else:
+            # manual implementation of attention
+            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+            att = F.softmax(att, dim=-1)
+            att = self.attn_dropout(att)
+            y = att @ v  # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+        y = (
+            y.transpose(1, 2).contiguous().view(B, T, C)
+        )  # re-assemble all head outputs side by side
+
+        # output projection
+        y = self.resid_dropout(self.c_proj(y))
+        return y
+
+
+class FineBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.ln_1 = nn.LayerNorm(config.n_embd)
+        self.attn = NonCausalSelfAttention(config)
+        self.ln_2 = nn.LayerNorm(config.n_embd)
+        self.mlp = MLP(config)
+
+    def forward(self, x):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class FineGPT(GPT):
+    def __init__(self, config):
+        super().__init__(config)
+        del self.lm_head
+        self.config = config
+        self.n_codes_total = config.n_codes_total
+        self.transformer = nn.ModuleDict(
+            dict(
+                wtes=nn.ModuleList(
+                    [
+                        nn.Embedding(config.input_vocab_size, config.n_embd)
+                        for _ in range(config.n_codes_total)
+                    ]
+                ),
+                wpe=nn.Embedding(config.block_size, config.n_embd),
+                drop=nn.Dropout(config.dropout),
+                h=nn.ModuleList([FineBlock(config) for _ in range(config.n_layer)]),
+                ln_f=nn.LayerNorm(config.n_embd),
+            )
+        )
+        self.lm_heads = nn.ModuleList(
+            [
+                nn.Linear(config.n_embd, config.output_vocab_size, bias=False)
+                for _ in range(config.n_codes_given, self.n_codes_total)
+            ]
+        )
+        for i in range(self.n_codes_total - config.n_codes_given):
+            self.transformer.wtes[i + 1].weight = self.lm_heads[i].weight
+
+    def forward(self, pred_idx, idx):
+        device = idx.device
+        b, t, codes = idx.size()
+        assert (
+            t <= self.config.block_size
+        ), f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"
+        assert pred_idx > 0, "cannot predict 0th codebook"
+        assert codes == self.n_codes_total, (b, t, codes)
+        pos = torch.arange(0, t, dtype=torch.long, device=device).unsqueeze(0)  # shape (1, t)
+
+        # forward the GPT model itself
+        tok_embs = [
+            wte(idx[:, :, i]).unsqueeze(-1) for i, wte in enumerate(self.transformer.wtes)
+        ]  # token embeddings of shape (b, t, n_embd)
+        tok_emb = torch.cat(tok_embs, dim=-1)
+        pos_emb = self.transformer.wpe(pos)  # position embeddings of shape (1, t, n_embd)
+        x = tok_emb[:, :, :, : pred_idx + 1].sum(dim=-1)
+        x = self.transformer.drop(x + pos_emb)
+        for block in self.transformer.h:
+            x = block(x)
+        x = self.transformer.ln_f(x)
+        logits = self.lm_heads[pred_idx - self.config.n_codes_given](x)
+        return logits
+
+    def get_num_params(self, non_embedding=True):
+        """
+        Return the number of parameters in the model.
+        For non-embedding count (default), the position embeddings get subtracted.
+        The token embeddings would too, except due to the parameter sharing these
+        params are actually used as weights in the final layer, so we include them.
+        """
+        n_params = sum(p.numel() for p in self.parameters())
+        if non_embedding:
+            for wte in self.transformer.wtes:
+                n_params -= wte.weight.numel()
+            n_params -= self.transformer.wpe.weight.numel()
+        return n_params
+
+
+@dataclass
+class FineGPTConfig(GPTConfig):
+    n_codes_total: int = 8
+    n_codes_given: int = 1

bark/settings.py

@@ -0,0 +1,7 @@
+import os
+
+def initenv(args):
+    os.environ['SUNO_USE_SMALL_MODELS'] = str("-smallmodels" in args)
+    os.environ['BARK_FORCE_CPU'] = str("-forcecpu" in args)
+    os.environ['SUNO_ENABLE_MPS'] = str("-enablemps" in args)
+    os.environ['SUNO_OFFLOAD_CPU'] = str("-offloadcpu" in args)

cloning/clonevoice.py

@@ -0,0 +1,68 @@
+from bark.generation import load_codec_model, generate_text_semantic, grab_best_device
+from encodec.utils import convert_audio
+from bark.hubert.hubert_manager import HuBERTManager
+from bark.hubert.pre_kmeans_hubert import CustomHubert
+from bark.hubert.customtokenizer import CustomTokenizer
+
+import torchaudio
+import torch
+import os
+import gradio
+
+
+def clone_voice(audio_filepath, dest_filename, progress=gradio.Progress(track_tqdm=True)):
+    # if len(text) < 1:
+    #    raise gradio.Error('No transcription text entered!')
+
+    use_gpu = not os.environ.get("BARK_FORCE_CPU", False)
+    progress(0, desc="Loading Codec")
+    model = load_codec_model(use_gpu=use_gpu)
+    
+    # From https://github.com/gitmylo/bark-voice-cloning-HuBERT-quantizer
+    hubert_manager = HuBERTManager()
+    hubert_manager.make_sure_hubert_installed()
+    hubert_manager.make_sure_tokenizer_installed()
+
+    # From https://github.com/gitmylo/bark-voice-cloning-HuBERT-quantizer 
+    # Load HuBERT for semantic tokens
+
+    # Load the HuBERT model
+    device = grab_best_device(use_gpu)
+    hubert_model = CustomHubert(checkpoint_path='./models/hubert/hubert.pt').to(device)
+
+    # Load the CustomTokenizer model
+    tokenizer = CustomTokenizer.load_from_checkpoint('./models/hubert/en_tokenizer.pth').to(device)  # change to the correct path 
+
+    progress(0.25, desc="Converting WAV")
+
+    # Load and pre-process the audio waveform
+    wav, sr = torchaudio.load(audio_filepath)
+    if wav.shape[0] == 2:  # Stereo to mono if needed
+        wav = wav.mean(0, keepdim=True)
+
+    wav = convert_audio(wav, sr, model.sample_rate, model.channels)
+    wav = wav.to(device)
+    progress(0.5, desc="Extracting codes")
+
+    semantic_vectors = hubert_model.forward(wav, input_sample_hz=model.sample_rate)
+    semantic_tokens = tokenizer.get_token(semantic_vectors)
+    
+    # Extract discrete codes from EnCodec
+    with torch.no_grad():
+        encoded_frames = model.encode(wav.unsqueeze(0))
+    codes = torch.cat([encoded[0] for encoded in encoded_frames], dim=-1).squeeze()  # [n_q, T]
+
+    # get seconds of audio
+    # seconds = wav.shape[-1] / model.sample_rate
+    # generate semantic tokens
+    # semantic_tokens = generate_text_semantic(text, max_gen_duration_s=seconds, top_k=50, top_p=.95, temp=0.7)
+
+    # move codes to cpu
+    codes = codes.cpu().numpy()
+    # move semantic tokens to cpu
+    semantic_tokens = semantic_tokens.cpu().numpy()
+
+    import numpy as np
+    output_path = dest_filename + '.npz'
+    np.savez(output_path, fine_prompt=codes, coarse_prompt=codes[:2, :], semantic_prompt=semantic_tokens)
+    return ["Finished", output_path]

config.yaml

@@ -0,0 +1,8 @@
+input_text_desired_length: 110
+input_text_max_length: 170
+selected_theme: JohnSmith9982/small_and_pretty
+server_name: ''
+server_port: 0
+server_share: false
+silence_between_sentences: 250
+silence_between_speakers: 500

pyproject.toml

@@ -0,0 +1,60 @@
+[build-system]
+requires = ["setuptools"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "bark-ui-enhanced"
+version = "0.7.0"
+description = "Bark text to audio model with addition features and a Web UI"
+readme = "README.md"
+requires-python = ">=3.8"
+authors =  [
+    {name = "Suno Inc (original Bark)", email = "hello@suno.ai"},
+    {name = "Count Floyd"},
+]
+# MIT License
+license = {file = "LICENSE"}
+
+dependencies = [
+    "boto3",
+    "encodec",
+    "funcy",
+    "huggingface-hub>=0.14.1",
+    "numpy",
+    "scipy",
+    "tokenizers",
+    "torch",
+    "tqdm",
+    "transformers",
+]
+
+[project.urls]
+source = "https://github.com/C0untFloyd/bark-gui"
+
+[project.optional-dependencies]
+dev = [
+    "bandit",
+    "black",
+    "codecov",
+    "flake8",
+    "hypothesis>=6.14,<7",
+    "isort>=5.0.0,<6",
+    "jupyter",
+    "mypy",
+    "nbconvert",
+    "nbformat",
+    "pydocstyle",
+    "pylint",
+    "pytest",
+    "pytest-cov",
+]
+
+[tool.setuptools]
+packages = ["bark"]
+
+[tool.setuptools.package-data]
+bark = ["assets/prompts/*.npz", "assets/prompts/v2/*.npz"]
+
+
+[tool.black]
+line-length = 100

requirements.txt

@@ -0,0 +1,13 @@
+fairseq; platform_system != "Windows"
+fairseq@https://github.com/Sharrnah/fairseq/releases/download/v0.12.4/fairseq-0.12.4-cp310-cp310-win_amd64.whl; platform_system == "Windows"
+audiolm-pytorch
+gradio
+funcy
+linkify
+mutagen
+pytorch_seed
+pyyaml
+sentencepiece
+soundfile; platform_system == "Windows"
+sox; platform_system != "Windows"
+transformers

setup.py

@@ -0,0 +1,3 @@
+from setuptools import setup
+
+setup()

swap_voice.py

@@ -0,0 +1,62 @@
+from bark.generation import load_codec_model, generate_text_semantic, grab_best_device
+from bark import SAMPLE_RATE
+from encodec.utils import convert_audio
+from bark.hubert.hubert_manager import HuBERTManager
+from bark.hubert.pre_kmeans_hubert import CustomHubert
+from bark.hubert.customtokenizer import CustomTokenizer
+from bark.api import semantic_to_waveform
+from scipy.io.wavfile import write as write_wav
+from util.helper import create_filename
+from util.settings import Settings
+
+
+import torchaudio
+import torch
+import os
+import gradio
+
+def swap_voice_from_audio(swap_audio_filename, selected_speaker, tokenizer_lang, seed, batchcount, progress=gradio.Progress(track_tqdm=True)):
+    use_gpu = not os.environ.get("BARK_FORCE_CPU", False)
+    progress(0, desc="Loading Codec")
+    
+    # From https://github.com/gitmylo/bark-voice-cloning-HuBERT-quantizer
+    hubert_manager = HuBERTManager()
+    hubert_manager.make_sure_hubert_installed()
+    hubert_manager.make_sure_tokenizer_installed(tokenizer_lang=tokenizer_lang)
+
+    # From https://github.com/gitmylo/bark-voice-cloning-HuBERT-quantizer 
+    # Load HuBERT for semantic tokens
+
+    # Load the HuBERT model
+    device = grab_best_device(use_gpu)
+    hubert_model = CustomHubert(checkpoint_path='./models/hubert/hubert.pt').to(device)
+    model = load_codec_model(use_gpu=use_gpu)
+
+    # Load the CustomTokenizer model
+    tokenizer = CustomTokenizer.load_from_checkpoint(f'./models/hubert/{tokenizer_lang}_tokenizer.pth').to(device)  # Automatically uses the right layers
+
+    progress(0.25, desc="Converting WAV")
+
+    # Load and pre-process the audio waveform
+    wav, sr = torchaudio.load(swap_audio_filename)
+    if wav.shape[0] == 2:  # Stereo to mono if needed
+        wav = wav.mean(0, keepdim=True)
+
+    wav = convert_audio(wav, sr, model.sample_rate, model.channels)
+    wav = wav.to(device)
+    semantic_vectors = hubert_model.forward(wav, input_sample_hz=model.sample_rate)
+    semantic_tokens = tokenizer.get_token(semantic_vectors)
+
+    audio = semantic_to_waveform(
+        semantic_tokens,
+        history_prompt=selected_speaker,
+        temp=0.7,
+        silent=False,
+        output_full=False)
+
+    settings = Settings('config.yaml')
+
+    result = create_filename(settings.output_folder_path, None, "swapvoice",".wav")
+    write_wav(result, SAMPLE_RATE, audio)
+    return result
+

training/data.py

@@ -0,0 +1,52 @@
+import random
+import requests
+import os, glob
+
+# english literature
+books = [
+     'https://www.gutenberg.org/cache/epub/1513/pg1513.txt',
+     'https://www.gutenberg.org/files/2701/2701-0.txt',
+     'https://www.gutenberg.org/cache/epub/84/pg84.txt',
+     'https://www.gutenberg.org/cache/epub/2641/pg2641.txt',
+     'https://www.gutenberg.org/cache/epub/1342/pg1342.txt',
+     'https://www.gutenberg.org/cache/epub/100/pg100.txt'
+ ]
+
+#default english
+# allowed_chars = ' abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789!@#$%^&*()-_+=\"\':;[]{}/<>,.`~\n\\'
+
+#german
+allowed_chars = ' aäbcdefghijklmnoöpqrsßtuüvwxyzABCDEFGHIJKLMNOÖPQRSTUÜVWXYZ0123456789!@#$%^&*()-_+=\"\':;[]{}/<>,.`~\n\\'
+
+
+def download_book(book):
+    return requests.get(book).content.decode('utf-8')
+
+
+def filter_data(data):
+    print('Filtering data')
+    return ''.join([char for char in data if char in allowed_chars])
+
+
+def load_books(fromfolder=False):
+    text_data = []
+    if fromfolder:
+        current_working_directory = os.getcwd()
+        print(current_working_directory)
+        path = 'text'
+        for filename in glob.glob(os.path.join(path, '*.txt')):
+            with open(os.path.join(os.getcwd(), filename), 'r') as f: # open in readonly mode
+                print(f'Loading {filename}')
+                text_data.append(filter_data(str(f.read())))
+    else:
+        print(f'Loading {len(books)} books into ram')
+        for book in books:
+            text_data.append(filter_data(str(download_book(book))))
+    print('Loaded books')
+    return ' '.join(text_data)
+
+
+def random_split_chunk(data, size=14):
+    data = data.split(' ')
+    index = random.randrange(0, len(data))
+    return ' '.join(data[index:index+size])

training/train.py

@@ -0,0 +1,47 @@
+import os
+import fnmatch
+import shutil
+
+import numpy
+import torchaudio
+import gradio
+
+from bark.hubert.pre_kmeans_hubert import CustomHubert
+from bark.hubert.customtokenizer import auto_train
+from tqdm.auto import tqdm
+
+
+def training_prepare_files(path, model,progress=gradio.Progress(track_tqdm=True)):
+
+    semanticsfolder = "./training/data/output"
+    wavfolder = "./training/data/output_wav"
+    ready = os.path.join(path, 'ready')
+
+    testfiles = fnmatch.filter(os.listdir(ready), '*.npy')
+    if(len(testfiles) < 1):
+        # prepare and copy for training
+        hubert_model = CustomHubert(checkpoint_path=model)
+
+        wavfiles = fnmatch.filter(os.listdir(wavfolder), '*.wav')
+        for i, f in tqdm(enumerate(wavfiles), total=len(wavfiles)):
+            semaname = '.'.join(f.split('.')[:-1])  # Cut off the extension
+            semaname = f'{semaname}.npy'
+            semafilename =  os.path.join(semanticsfolder, semaname)
+            if not os.path.isfile(semafilename):
+                print(f'Skipping {f} no semantics pair found!')
+                continue
+
+            print('Processing', f)
+            wav, sr = torchaudio.load(os.path.join(wavfolder, f))
+            if wav.shape[0] == 2:  # Stereo to mono if needed
+                wav = wav.mean(0, keepdim=True)
+            output = hubert_model.forward(wav, input_sample_hz=sr)
+            out_array = output.cpu().numpy()
+            fname = f'{i}_semantic_features.npy'
+            numpy.save(os.path.join(ready, fname), out_array)
+            fname = f'{i}_semantic.npy'
+            shutil.copy(semafilename, os.path.join(ready, fname))
+
+def train(path, save_every, max_epochs):
+    auto_train(path, save_epochs=save_every)
+

training/training_prepare.py

@@ -0,0 +1,73 @@
+import random
+import uuid
+import numpy
+import os
+import random
+import fnmatch
+
+from tqdm.auto import tqdm
+from scipy.io import wavfile
+
+from bark.generation import load_model, SAMPLE_RATE
+from bark.api import semantic_to_waveform
+
+from bark import text_to_semantic
+from bark.generation import load_model
+
+from training.data import load_books, random_split_chunk
+
+output = 'training/data/output'
+output_wav = 'training/data/output_wav'
+
+
+def prepare_semantics_from_text(num_generations):
+    loaded_data = load_books(True)
+
+    print('Loading semantics model')
+    load_model(use_gpu=True, use_small=False, force_reload=False, model_type='text')
+
+    if not os.path.isdir(output):
+        os.mkdir(output)
+
+    loop = 1
+    while 1:
+        filename = uuid.uuid4().hex + '.npy'
+        file_name = os.path.join(output, filename)
+        text = ''
+        while not len(text) > 0:
+            text = random_split_chunk(loaded_data)  # Obtain a short chunk of text
+            text = text.strip()
+        print(f'{loop} Generating semantics for text:', text)
+        loop+=1 
+        semantics = text_to_semantic(text, temp=round(random.uniform(0.6, 0.8), ndigits=2))
+        numpy.save(file_name, semantics)
+
+
+def prepare_wavs_from_semantics():
+    if not os.path.isdir(output):
+        raise Exception('No \'output\' folder, make sure you run create_data.py first!')
+    if not os.path.isdir(output_wav):
+        os.mkdir(output_wav)
+
+    print('Loading coarse model')
+    load_model(use_gpu=True, use_small=False, force_reload=False, model_type='coarse')
+    print('Loading fine model')
+    load_model(use_gpu=True, use_small=False, force_reload=False, model_type='fine')
+
+    files = fnmatch.filter(os.listdir(output), '*.npy')
+    current = 1
+    total = len(files)
+
+    for i, f in tqdm(enumerate(files), total=len(files)):
+        real_name = '.'.join(f.split('.')[:-1])  # Cut off the extension
+        file_name = os.path.join(output, f)
+        out_file = os.path.join(output_wav, f'{real_name}.wav')
+        if not os.path.isfile(out_file) and os.path.isfile(file_name):  # Don't process files that have already been processed, to be able to continue previous generations
+            print(f'Processing ({i+1}/{total}) -> {f}')
+            wav = semantic_to_waveform(numpy.load(file_name), temp=round(random.uniform(0.6, 0.8), ndigits=2))
+            # Change to PCM16
+            # wav = (wav * 32767).astype(np.int16)
+            wavfile.write(out_file, SAMPLE_RATE, wav)
+
+    print('Done!')
+

util/helper.py

@@ -0,0 +1,35 @@
+import os
+from datetime import datetime
+from mutagen.wave import WAVE
+from mutagen.id3._frames import *
+
+def create_filename(path, seed, name, extension):
+    now = datetime.now()
+    date_str =now.strftime("%m-%d-%Y")
+    outputs_folder = os.path.join(os.getcwd(), path)
+    if not os.path.exists(outputs_folder):
+        os.makedirs(outputs_folder)
+
+    sub_folder = os.path.join(outputs_folder, date_str)
+    if not os.path.exists(sub_folder):
+        os.makedirs(sub_folder)
+
+    time_str = now.strftime("%H-%M-%S")
+    if seed == None:
+        file_name = f"{name}_{time_str}{extension}"
+    else:
+        file_name = f"{name}_{time_str}_s{seed}{extension}"
+    return os.path.join(sub_folder, file_name)
+
+
+def add_id3_tag(filename, text, speakername, seed):
+    audio = WAVE(filename)
+    if speakername == None:
+        speakername = "Unconditional"
+
+    # write id3 tag with text truncated to 60 chars, as a precaution...
+    audio["TIT2"] = TIT2(encoding=3, text=text[:60])
+    audio["TPE1"] = TPE1(encoding=3, text=f"Voice {speakername} using Seed={seed}")
+    audio["TPUB"] = TPUB(encoding=3, text="Bark by Suno AI")
+    audio["COMMENT"] = COMM(encoding=3, text="Generated with Bark GUI - Text-Prompted Generative Audio Model. Visit https://github.com/C0untFloyd/bark-gui")
+    audio.save()

util/parseinput.py

@@ -0,0 +1,129 @@
+import re
+import xml.etree.ElementTree as ET
+from xml.sax import saxutils
+#import nltk
+
+# Chunked generation originally from https://github.com/serp-ai/bark-with-voice-clone
+def split_and_recombine_text(text, desired_length=100, max_length=150):
+    # return nltk.sent_tokenize(text)
+
+    # from https://github.com/neonbjb/tortoise-tts
+    """Split text it into chunks of a desired length trying to keep sentences intact."""
+    # normalize text, remove redundant whitespace and convert non-ascii quotes to ascii
+    text = re.sub(r"\n\n+", "\n", text)
+    text = re.sub(r"\s+", " ", text)
+    text = re.sub(r"[“”]", '"', text)
+
+    rv = []
+    in_quote = False
+    current = ""
+    split_pos = []
+    pos = -1
+    end_pos = len(text) - 1
+
+    def seek(delta):
+        nonlocal pos, in_quote, current
+        is_neg = delta < 0
+        for _ in range(abs(delta)):
+            if is_neg:
+                pos -= 1
+                current = current[:-1]
+            else:
+                pos += 1
+                current += text[pos]
+            if text[pos] == '"':
+                in_quote = not in_quote
+        return text[pos]
+
+    def peek(delta):
+        p = pos + delta
+        return text[p] if p < end_pos and p >= 0 else ""
+
+    def commit():
+        nonlocal rv, current, split_pos
+        rv.append(current)
+        current = ""
+        split_pos = []
+
+    while pos < end_pos:
+        c = seek(1)
+        # do we need to force a split?
+        if len(current) >= max_length:
+            if len(split_pos) > 0 and len(current) > (desired_length / 2):
+                # we have at least one sentence and we are over half the desired length, seek back to the last split
+                d = pos - split_pos[-1]
+                seek(-d)
+            else:
+                # no full sentences, seek back until we are not in the middle of a word and split there
+                while c not in "!?.,\n " and pos > 0 and len(current) > desired_length:
+                    c = seek(-1)
+            commit()
+        # check for sentence boundaries
+        elif not in_quote and (c in "!?]\n" or (c == "." and peek(1) in "\n ")):
+            # seek forward if we have consecutive boundary markers but still within the max length
+            while (
+                pos < len(text) - 1 and len(current) < max_length and peek(1) in "!?.]"
+            ):
+                c = seek(1)
+            split_pos.append(pos)
+            if len(current) >= desired_length:
+                commit()
+        # treat end of quote as a boundary if its followed by a space or newline
+        elif in_quote and peek(1) == '"' and peek(2) in "\n ":
+            seek(2)
+            split_pos.append(pos)
+    rv.append(current)
+
+    # clean up, remove lines with only whitespace or punctuation
+    rv = [s.strip() for s in rv]
+    rv = [s for s in rv if len(s) > 0 and not re.match(r"^[\s\.,;:!?]*$", s)]
+
+    return rv
+
+def is_ssml(value):
+    try:
+        ET.fromstring(value)
+    except ET.ParseError:
+        return False
+    return True
+
+def build_ssml(rawtext, selected_voice):
+    texts = rawtext.split("\n")
+    joinedparts = ""
+    for textpart in texts:
+        textpart = textpart.strip()
+        if len(textpart) < 1:
+            continue
+        joinedparts = joinedparts + f"\n<voice name=\"{selected_voice}\">{saxutils.escape(textpart)}</voice>"
+    ssml = f"""<?xml version="1.0"?>
+<speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis"
+         xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
+         xsi:schemaLocation="http://www.w3.org/2001/10/synthesis
+                   http://www.w3.org/TR/speech-synthesis/synthesis.xsd"
+         xml:lang="en-US">
+         {joinedparts}
+</speak>
+    """
+    return ssml
+
+def create_clips_from_ssml(ssmlinput):
+    # Parse the XML
+    tree = ET.ElementTree(ET.fromstring(ssmlinput))
+    root = tree.getroot()
+
+    # Create an empty list
+    voice_list = []
+
+    # Loop through all voice tags
+    for voice in root.iter('{http://www.w3.org/2001/10/synthesis}voice'):
+        # Extract the voice name attribute and the content text
+        voice_name = voice.attrib['name']
+        voice_content = voice.text.strip() if voice.text else ''
+        if(len(voice_content) > 0):
+            parts = split_and_recombine_text(voice_content)
+            for p in parts:
+                if(len(p) > 1):
+                    # add to tuple list
+                    voice_list.append((voice_name, p))
+    return voice_list
+