import gradio as gr import ljspeechimportable import torch import numpy as np import styletts2importable import re import phonemizer 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 theme = gr.themes.Base( font=[gr.themes.GoogleFont('Libre Franklin'), gr.themes.GoogleFont('Public Sans'), 'system-ui', 'sans-serif'], ) voicelist = ['f-us-1', 'f-us-2', 'f-us-3', 'f-us-4', 'm-us-1', 'm-us-2', 'm-us-3', 'm-us-4'] voices = {} global_phonemizer = phonemizer.backend.EspeakBackend(language='en-us', preserve_punctuation=True, with_stress=True) for v in voicelist: voices[v] = styletts2importable.compute_style(f'voices/{v}.wav') 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 = [] for t in progress.tqdm(texts): 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 ljsynthesize(text, steps, progress=gr.Progress()): noise = torch.randn(1,1,256).to('cuda' if torch.cuda.is_available() else 'cpu') 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 libritts: # 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) 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 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.DuplicateButton("Duplicate Space") gr.TabbedInterface([libritts, lj], ['Multi-Voice', 'LJSpeech']) gr.Markdown(""" Original 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)