Partnership in Asia

README.md

@@ -8,7 +8,6 @@ sdk_version: 4.5.0
 app_file: app.py
 pinned: true
 license: agpl-3.0
-short_description: Efficient, fast, and natural text to speech with StyleTTS 2!
 ---
 
 License: AGPLv3

app.py

@@ -1,23 +1,23 @@
-INTROTXT = """# StyleTTS 2
+# # # # # # # # # # # # # # # # # # # # # # # # # # # #
+#                                                     #
+#                   StyleTTS 2 Demo                   #
+#                                                     #
+#                                                     #
+# Copyright (c) 2023 mrfakename. All rights reserved. #
+#                                                     #
+#   License : AGPL v3                                 #
+#   Version : 2.0                                     #
+#   Support : https://github.com/neuralvox/styletts2  #
+#                                                     #
+# # # # # # # # # # # # # # # # # # # # # # # # # # # #
 
-[Paper](https://arxiv.org/abs/2306.07691) - [Samples](https://styletts2.github.io/) - [Code](https://github.com/yl4579/StyleTTS2) - [Discord](https://discord.gg/ha8sxdG2K4)
 
-A free demo of StyleTTS 2. **I am not affiliated with the StyleTTS 2 Authors.**
-
-**Before using this demo, you agree to inform the listeners that the speech samples are synthesized by the pre-trained models, unless you have the permission to use the voice you synthesize. That is, you agree to only use voices whose speakers grant the permission to have their voice cloned, either directly or by license before making synthesized voices public, or you have to publicly announce that these voices are synthesized if you do not have the permission to use these voices.**
-
-Is there a long queue on this space? Duplicate it and add a more powerful GPU to skip the wait! **Note: Thank you to Hugging Face for their generous GPU grant program!**
-
-**NOTE: StyleTTS 2 does better on longer texts.** For example, making it say "hi" will produce a lower-quality result than making it say a longer phrase.
-
-**NOTE: StyleTTS 2 is _currently_ English-only. Join the Discord for updates on multilingual training.**
-"""
 import gradio as gr
 import styletts2importable
 import ljspeechimportable
 import torch
 import os
-from txtsplit import txtsplit
+from tortoise.utils.text import split_and_recombine_text
 import numpy as np
 import pickle
 theme = gr.themes.Base(
@@ -43,20 +43,15 @@ for v in voicelist:
 #     v = voice.lower()
 #     # return (24000, styletts2importable.inference(text, voices[v], alpha=0.3, beta=0.7, diffusion_steps=7, embedding_scale=1))
 #     return (24000, styletts2importable.inference(text, voices[v], alpha=0.3, beta=0.7, diffusion_steps=multispeakersteps, embedding_scale=1))
-if not torch.cuda.is_available(): INTROTXT += "\n\n### You are on a CPU-only system, inference will be much slower.\n\nYou can use the [online demo](https://huggingface.co/spaces/styletts2/styletts2) for fast inference."
 def synthesize(text, voice, lngsteps, password, progress=gr.Progress()):
     if text.strip() == "":
         raise gr.Error("You must enter some text")
     if len(text) > 50000:
         raise gr.Error("Text must be <50k characters")
-    print("*** saying ***")
-    print(text)
-    print("*** end ***")
-    texts = txtsplit(text)
+    texts = split_and_recombine_text(text)
     v = voice.lower()
     audios = []
     for t in progress.tqdm(texts):
-        print(t)
         audios.append(styletts2importable.inference(t, voices[v], alpha=0.3, beta=0.7, diffusion_steps=lngsteps, embedding_scale=1))
     return (24000, np.concatenate(audios))
 # def longsynthesize(text, voice, lngsteps, password, progress=gr.Progress()):
@@ -75,7 +70,7 @@ def synthesize(text, voice, lngsteps, password, progress=gr.Progress()):
 #         return (24000, np.concatenate(audios))
 #     else:
 #         raise gr.Error('Wrong access code')
-def clsynthesize(text, voice, vcsteps, embscale, alpha, beta, progress=gr.Progress()):
+def clsynthesize(text, voice, vcsteps, progress=gr.Progress()):
     # if text.strip() == "":
     #     raise gr.Error("You must enter some text")
     # # if global_phonemizer.phonemize([text]) > 300:
@@ -87,19 +82,10 @@ def clsynthesize(text, voice, vcsteps, embscale, alpha, beta, progress=gr.Progre
         raise gr.Error("You must enter some text")
     if len(text) > 50000:
         raise gr.Error("Text must be <50k characters")
-    if embscale > 1.3 and len(text) < 20:
-        gr.Warning("WARNING: You entered short text, you may get static!")
-    print("*** saying ***")
-    print(text)
-    print("*** end ***")
-    texts = txtsplit(text)
+    texts = split_and_recombine_text(text)
     audios = []
-    # vs = styletts2importable.compute_style(voice)
-    vs = styletts2importable.compute_style(voice)
-    # print(vs)
     for t in progress.tqdm(texts):
-        audios.append(styletts2importable.inference(t, vs, alpha=alpha, beta=beta, diffusion_steps=vcsteps, embedding_scale=embscale))
-        # audios.append(styletts2importable.inference(t, vs, diffusion_steps=10, alpha=0.3, beta=0.7, embedding_scale=5))
+        audios.append(styletts2importable.inference(t, styletts2importable.compute_style(voice), alpha=0.3, beta=0.7, diffusion_steps=vcsteps, embedding_scale=1))
     return (24000, np.concatenate(audios))
 def ljsynthesize(text, steps, progress=gr.Progress()):
     # if text.strip() == "":
@@ -113,17 +99,14 @@ def ljsynthesize(text, steps, progress=gr.Progress()):
         raise gr.Error("You must enter some text")
     if len(text) > 150000:
         raise gr.Error("Text must be <150k characters")
-    print("*** saying ***")
-    print(text)
-    print("*** end ***")
-    texts = txtsplit(text)
+    texts = split_and_recombine_text(text)
     audios = []
     for t in progress.tqdm(texts):
         audios.append(ljspeechimportable.inference(t, noise, diffusion_steps=steps, embedding_scale=1))
     return (24000, np.concatenate(audios))
 
 
-with gr.Blocks() as vctk:
+with gr.Blocks() as vctk: # just realized it isn't vctk but libritts but i'm too lazy to change it rn
     with gr.Row():
         with gr.Column(scale=1):
             inp = gr.Textbox(label="Text", info="What would you like StyleTTS 2 to read? It works better on full sentences.", interactive=True)
@@ -132,21 +115,18 @@ with gr.Blocks() as vctk:
             # use_gruut = gr.Checkbox(label="Use alternate phonemizer (Gruut) - Experimental")
         with gr.Column(scale=1):
             btn = gr.Button("Synthesize", variant="primary")
-            audio = gr.Audio(interactive=False, label="Synthesized Audio", waveform_options={'waveform_progress_color': '#3C82F6'})
+            audio = gr.Audio(interactive=False, label="Synthesized Audio")
             btn.click(synthesize, inputs=[inp, voice, multispeakersteps], outputs=[audio], concurrency_limit=4)
 with gr.Blocks() as clone:
     with gr.Row():
         with gr.Column(scale=1):
             clinp = gr.Textbox(label="Text", info="What would you like StyleTTS 2 to read? It works better on full sentences.", interactive=True)
-            clvoice = gr.Audio(label="Voice", interactive=True, type='filepath', max_length=300, waveform_options={'waveform_progress_color': '#3C82F6'})
+            clvoice = gr.Audio(label="Voice", interactive=True, type='filepath', max_length=300)
             vcsteps = gr.Slider(minimum=3, maximum=20, value=20, step=1, label="Diffusion Steps", info="Theoretically, higher should be better quality but slower, but we cannot notice a difference. Try with lower steps first - it is faster", interactive=True)
-            embscale = gr.Slider(minimum=1, maximum=10, value=1, step=0.1, label="Embedding Scale (READ WARNING BELOW)", info="Defaults to 1. WARNING: If you set this too high and generate text that's too short you will get static!", interactive=True)
-            alpha = gr.Slider(minimum=0, maximum=1, value=0.3, step=0.1, label="Alpha", info="Defaults to 0.3", interactive=True)
-            beta = gr.Slider(minimum=0, maximum=1, value=0.7, step=0.1, label="Beta", info="Defaults to 0.7", interactive=True)
         with gr.Column(scale=1):
             clbtn = gr.Button("Synthesize", variant="primary")
-            claudio = gr.Audio(interactive=False, label="Synthesized Audio", waveform_options={'waveform_progress_color': '#3C82F6'})
-            clbtn.click(clsynthesize, inputs=[clinp, clvoice, vcsteps, embscale, alpha, beta], outputs=[claudio], concurrency_limit=4)
+            claudio = gr.Audio(interactive=False, label="Synthesized Audio")
+            clbtn.click(clsynthesize, inputs=[clinp, clvoice, vcsteps], outputs=[claudio], concurrency_limit=4)
 # with gr.Blocks() as longText:
 #     with gr.Row():
 #         with gr.Column(scale=1):
@@ -165,10 +145,22 @@ with gr.Blocks() as lj:
             ljsteps = gr.Slider(minimum=3, maximum=20, value=3, step=1, label="Diffusion Steps", info="Theoretically, higher should be better quality but slower, but we cannot notice a difference. Try with lower steps first - it is faster", interactive=True)
         with gr.Column(scale=1):
             ljbtn = gr.Button("Synthesize", variant="primary")
-            ljaudio = gr.Audio(interactive=False, label="Synthesized Audio", waveform_options={'waveform_progress_color': '#3C82F6'})
+            ljaudio = gr.Audio(interactive=False, label="Synthesized Audio")
             ljbtn.click(ljsynthesize, inputs=[ljinp, ljsteps], outputs=[ljaudio], concurrency_limit=4)
 with gr.Blocks(title="StyleTTS 2", css="footer{display:none !important}", theme=theme) as demo:
-    gr.Markdown(INTROTXT)
+    gr.Markdown("""# StyleTTS 2
+
+[Paper](https://arxiv.org/abs/2306.07691) - [Samples](https://styletts2.github.io/) - [Code](https://github.com/yl4579/StyleTTS2)
+
+A free demo of StyleTTS 2. **I am not affiliated with the StyleTTS 2 Authors.**
+
+#### Help this space get to the top of HF's trending list! Please give this space a Like!
+
+**Before using this demo, you agree to inform the listeners that the speech samples are synthesized by the pre-trained models, unless you have the permission to use the voice you synthesize. That is, you agree to only use voices whose speakers grant the permission to have their voice cloned, either directly or by license before making synthesized voices public, or you have to publicly announce that these voices are synthesized if you do not have the permission to use these voices.**
+
+Is there a long queue on this space? Duplicate it and add a more powerful GPU to skip the wait! **Note: Thank you to Hugging Face for their generous GPU grant program!**
+
+**NOTE: StyleTTS 2 does better on longer texts.** For example, making it say "hi" will produce a lower-quality result than making it say a longer phrase.""")
     gr.DuplicateButton("Duplicate Space")
     # gr.TabbedInterface([vctk, clone, lj, longText], ['Multi-Voice', 'Voice Cloning', 'LJSpeech', 'Long Text [Beta]'])
     gr.TabbedInterface([vctk, clone, lj], ['Multi-Voice', 'Voice Cloning', 'LJSpeech', 'Long Text [Beta]'])

requirements.txt

@@ -1,24 +1,23 @@
-SoundFile
-torchaudio
-munch
-torch
-pydub
-pyyaml
-librosa
-nltk
-matplotlib
-accelerate
-transformers
-einops
-einops-exts
-tqdm
-typing
-typing-extensions
-git+https://github.com/resemble-ai/monotonic_align.git
-scipy
-phonemizer
-cached-path
-gradio
-gruut
-#tortoise-tts
-txtsplit
+SoundFile
+torchaudio
+munch
+torch
+pydub
+pyyaml
+librosa
+nltk
+matplotlib
+accelerate
+transformers
+einops
+einops-exts
+tqdm
+typing
+typing-extensions
+git+https://github.com/resemble-ai/monotonic_align.git
+scipy
+phonemizer
+cached-path
+gradio
+gruut
+tortoise-tts