Update app.py

app.py

@@ -1,25 +1,8 @@
-# # # # # # # # # # # # # # # # # # # # # # # # # # # #
-#                                                     #
-#                   StyleTTS 2 Demo                   #
-#                                                     #
-#                                                     #
-# Copyright (c) 2023 mrfakename. All rights reserved. #
-#                                                     #
-#   License : AGPL v3                                 #
-#   Version : 2.0                                     #
-#   Support : https://github.com/neuralvox/styletts2  #
-#                                                     #
-# # # # # # # # # # # # # # # # # # # # # # # # # # # #
-
-
 import gradio as gr
 import styletts2importable
-import ljspeechimportable
-import torch
-import os
-from tortoise.utils.text import split_and_recombine_text
 import numpy as np
-import pickle
+import re
+
 theme = gr.themes.Base(
     font=[gr.themes.GoogleFont('Libre Franklin'), gr.themes.GoogleFont('Public Sans'), 'system-ui', 'sans-serif'],
 )
@@ -27,26 +10,83 @@ voicelist = ['f-us-1', 'f-us-2', 'f-us-3', 'f-us-4', 'm-us-1', 'm-us-2', 'm-us-3
 voices = {}
 import phonemizer
 global_phonemizer = phonemizer.backend.EspeakBackend(language='en-us', preserve_punctuation=True,  with_stress=True)
-# todo: cache computed style, load using pickle
-# if os.path.exists('voices.pkl'):
-    # with open('voices.pkl', 'rb') as f:
-        # voices = pickle.load(f)
-# else:
 for v in voicelist:
     voices[v] = styletts2importable.compute_style(f'voices/{v}.wav')
-# def synthesize(text, voice, multispeakersteps):
-#     if text.strip() == "":
-#         raise gr.Error("You must enter some text")
-#     # if len(global_phonemizer.phonemize([text])) > 300:
-#     if len(text) > 300:
-#         raise gr.Error("Text must be under 300 characters")
-#     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))
+def split_and_recombine_text(text, desired_length=200, max_length=400):
+    """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 synthesize(text, voice, lngsteps, password, progress=gr.Progress()):
     if text.strip() == "":
         raise gr.Error("You must enter some text")
-
+        
     texts = split_and_recombine_text(text)
     v = voice.lower()
     audios = []
@@ -54,94 +94,16 @@ def synthesize(text, voice, lngsteps, password, progress=gr.Progress()):
         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 clsynthesize(text, voice, vcsteps, progress=gr.Progress()):
-    if text.strip() == "":
-        raise gr.Error("You must enter some text")
-    texts = split_and_recombine_text(text)
-    audios = []
-    for t in progress.tqdm(texts):
-        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()):
-    noise = torch.randn(1,1,256).to('cuda' if torch.cuda.is_available() else 'cpu')
-    # return (24000, ljspeechimportable.inference(text, noise, diffusion_steps=7, embedding_scale=1))
-    if text.strip() == "":
-        raise gr.Error("You must enter some 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: # just realized it isn't vctk but libritts but i'm too lazy to change it rn
+with gr.Blocks() as demo:
     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)
             voice = gr.Dropdown(voicelist, label="Voice", info="Select a default voice.", value='m-us-2', interactive=True)
             multispeakersteps = gr.Slider(minimum=3, maximum=15, 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)
-            # 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")
             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)
-            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)
-        with gr.Column(scale=1):
-            clbtn = gr.Button("Synthesize", variant="primary")
-            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):
-#             lnginp = gr.Textbox(label="Text", info="What would you like StyleTTS 2 to read? It works better on full sentences.", interactive=True)
-#             lngvoice = gr.Dropdown(voicelist, label="Voice", info="Select a default voice.", value='m-us-1', interactive=True)
-#             lngsteps = gr.Slider(minimum=5, maximum=25, value=10, step=1, label="Diffusion Steps", info="Higher = better quality, but slower", interactive=True)
-#             lngpwd = gr.Textbox(label="Access code", info="This feature is in beta. You need an access code to use it as it uses more resources and we would like to prevent abuse")
-#         with gr.Column(scale=1):
-#             lngbtn = gr.Button("Synthesize", variant="primary")
-#             lngaudio = gr.Audio(interactive=False, label="Synthesized Audio")
-#             lngbtn.click(longsynthesize, inputs=[lnginp, lngvoice, lngsteps, lngpwd], outputs=[lngaudio], concurrency_limit=4)
-with gr.Blocks() as lj:
-    with gr.Row():
-        with gr.Column(scale=1):
-            ljinp = gr.Textbox(label="Text", info="What would you like StyleTTS 2 to read? It works better on full sentences.", interactive=True)
-            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")
-            ljbtn.click(ljsynthesize, inputs=[ljinp, ljsteps], outputs=[ljaudio], concurrency_limit=4)
-with gr.Blocks(title="StyleTTS 2", css="", theme=theme) as demo:
-    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]'])
-    gr.Markdown("""
-Demo by [mrfakename](https://twitter.com/realmrfakename). I am not affiliated with the StyleTTS 2 authors.
-
-Run this demo locally using Docker:
-
-```bash
-docker run -it -p 7860:7860 --platform=linux/amd64 --gpus all registry.hf.space/styletts2-styletts2:latest python app.py
-```
-""") # Please do not remove this line.
-if __name__ == "__main__":
-    # demo.queue(api_open=False, max_size=15).launch(show_api=False)
-    demo.queue(api_open=True, max_size=15).launch(show_api=True)
 
+demo.queue(api_open=True, max_size=15).launch(show_api=True)