fighter-programmer/voicegen
0
1import argparse2import json3import os4import re5import tempfile6import logging7 8logging.getLogger('numba').setLevel(logging.WARNING)9import librosa10import numpy as np11import torch12from torch import no_grad, LongTensor13import commons14import utils15import gradio as gr16import gradio.utils as gr_utils17import gradio.processing_utils as gr_processing_utils18import ONNXVITS_infer19import models20from text import text_to_sequence, _clean_text21from text.symbols import symbols22from mel_processing import spectrogram_torch23import psutil24from datetime import datetime25 26language_marks = {27 "Japanese": "",28 "日本語": "[JA]",29 "简体中文": "[ZH]",30 "English": "[EN]",31 "Mix": "",32}33 34limitation = os.getenv("SYSTEM") == "spaces" # limit text and audio length in huggingface spaces35 36 37def create_tts_fn(model, hps, speaker_ids):38 def tts_fn(text, speaker, language, speed, is_symbol):39 if limitation:40 text_len = len(re.sub("\[([A-Z]{2})\]", "", text))41 max_len = 15042 if is_symbol:43 max_len *= 344 if text_len > max_len:45 return "Error: Text is too long", None46 if language is not None:47 text = language_marks[language] + text + language_marks[language]48 speaker_id = speaker_ids[speaker]49 stn_tst = get_text(text, hps, is_symbol)50 with no_grad():51 x_tst = stn_tst.unsqueeze(0)52 x_tst_lengths = LongTensor([stn_tst.size(0)])53 sid = LongTensor([speaker_id])54 audio = model.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=.667, noise_scale_w=0.8,55 length_scale=1.0 / speed)[0][0, 0].data.cpu().float().numpy()56 del stn_tst, x_tst, x_tst_lengths, sid57 return "Success", (hps.data.sampling_rate, audio)58 59 return tts_fn60 61 62def create_vc_fn(model, hps, speaker_ids):63 def vc_fn(original_speaker, target_speaker, input_audio):64 if input_audio is None:65 return "You need to upload an audio", None66 sampling_rate, audio = input_audio67 duration = audio.shape[0] / sampling_rate68 if limitation and duration > 30:69 return "Error: Audio is too long", None70 original_speaker_id = speaker_ids[original_speaker]71 target_speaker_id = speaker_ids[target_speaker]72 73 audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)74 if len(audio.shape) > 1:75 audio = librosa.to_mono(audio.transpose(1, 0))76 if sampling_rate != hps.data.sampling_rate:77 audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=hps.data.sampling_rate)78 with no_grad():79 y = torch.FloatTensor(audio)80 y = y.unsqueeze(0)81 spec = spectrogram_torch(y, hps.data.filter_length,82 hps.data.sampling_rate, hps.data.hop_length, hps.data.win_length,83 center=False)84 spec_lengths = LongTensor([spec.size(-1)])85 sid_src = LongTensor([original_speaker_id])86 sid_tgt = LongTensor([target_speaker_id])87 audio = model.voice_conversion(spec, spec_lengths, sid_src=sid_src, sid_tgt=sid_tgt)[0][88 0, 0].data.cpu().float().numpy()89 del y, spec, spec_lengths, sid_src, sid_tgt90 return "Success", (hps.data.sampling_rate, audio)91 92 return vc_fn93 94 95def get_text(text, hps, is_symbol):96 text_norm = text_to_sequence(text, hps.symbols, [] if is_symbol else hps.data.text_cleaners)97 if hps.data.add_blank:98 text_norm = commons.intersperse(text_norm, 0)99 text_norm = LongTensor(text_norm)100 return text_norm101 102 103def create_to_symbol_fn(hps):104 def to_symbol_fn(is_symbol_input, input_text, temp_text):105 return (_clean_text(input_text, hps.data.text_cleaners), input_text) if is_symbol_input \106 else (temp_text, temp_text)107 108 return to_symbol_fn109 110 111models_tts = []112models_vc = []113models_info = [114 {115 "title": "Trilingual",116 "languages": ['日本語', '简体中文', 'English', 'Mix'],117 "description": """118 This model is trained on a mix up of Umamusume, Genshin Impact, Sanoba Witch & VCTK voice data to learn multilanguage.119 All characters can speak English, Chinese & Japanese.\n\n120 To mix multiple languages in a single sentence, wrap the corresponding part with language tokens121 ([JA] for Japanese, [ZH] for Chinese, [EN] for English), as shown in the examples.\n\n122 这个模型在赛马娘，原神，魔女的夜宴以及VCTK数据集上混合训练以学习多种语言。123 所有角色均可说中日英三语。\n\n124 若需要在同一个句子中混合多种语言，使用相应的语言标记包裹句子。125 （日语用[JA], 中文用[ZH], 英文用[EN]），参考Examples中的示例。126 """,127 "model_path": "./pretrained_models/G_trilingual.pth",128 "config_path": "./configs/uma_trilingual.json",129 "examples": [['你好，训练员先生，很高兴见到你。', '草上飞 Grass Wonder (Umamusume Pretty Derby)', '简体中文', 1, False],130 ['To be honest, I have no idea what to say as examples.', '派蒙 Paimon (Genshin Impact)', 'English',131 1, False],132 ['授業中に出しだら，学校生活終わるですわ。', '綾地 寧々 Ayachi Nene (Sanoba Witch)', '日本語', 1, False],133 ['[JA]こんにちわ。[JA][ZH]你好！[ZH][EN]Hello![EN]', '綾地 寧々 Ayachi Nene (Sanoba Witch)', 'Mix', 1, False]],134 "onnx_dir": "./ONNX_net/G_trilingual/"135 },136 {137 "title": "Japanese",138 "languages": ["Japanese"],139 "description": """140 This model contains 87 characters from Umamusume: Pretty Derby, Japanese only.\n\n141 这个模型包含赛马娘的所有87名角色，只能合成日语。142 """,143 "model_path": "./pretrained_models/G_jp.pth",144 "config_path": "./configs/uma87.json",145 "examples": [['お疲れ様です，トレーナーさん。', '无声铃鹿 Silence Suzuka (Umamusume Pretty Derby)', 'Japanese', 1, False],146 ['張り切っていこう！', '北部玄驹 Kitasan Black (Umamusume Pretty Derby)', 'Japanese', 1, False],147 ['何でこんなに慣れでんのよ，私のほが先に好きだっだのに。', '草上飞 Grass Wonder (Umamusume Pretty Derby)', 'Japanese', 1, False],148 ['授業中に出しだら，学校生活終わるですわ。', '目白麦昆 Mejiro Mcqueen (Umamusume Pretty Derby)', 'Japanese', 1, False],149 ['お帰りなさい，お兄様！', '米浴 Rice Shower (Umamusume Pretty Derby)', 'Japanese', 1, False],150 ['私の処女をもらっでください！', '米浴 Rice Shower (Umamusume Pretty Derby)', 'Japanese', 1, False]],151 "onnx_dir": "./ONNX_net/G_jp/"152 },153]154 155if __name__ == "__main__":156 parser = argparse.ArgumentParser()157 parser.add_argument("--share", action="store_true", default=False, help="share gradio app")158 args = parser.parse_args()159 for info in models_info:160 name = info['title']161 lang = info['languages']162 examples = info['examples']163 config_path = info['config_path']164 model_path = info['model_path']165 description = info['description']166 onnx_dir = info["onnx_dir"]167 hps = utils.get_hparams_from_file(config_path)168 model = ONNXVITS_infer.SynthesizerTrn(169 len(hps.symbols),170 hps.data.filter_length // 2 + 1,171 hps.train.segment_size // hps.data.hop_length,172 n_speakers=hps.data.n_speakers,173 ONNX_dir=onnx_dir,174 **hps.model)175 utils.load_checkpoint(model_path, model, None)176 model.eval()177 speaker_ids = hps.speakers178 speakers = list(hps.speakers.keys())179 models_tts.append((name, description, speakers, lang, examples,180 hps.symbols, create_tts_fn(model, hps, speaker_ids),181 create_to_symbol_fn(hps)))182 models_vc.append((name, description, speakers, create_vc_fn(model, hps, speaker_ids)))183 app = gr.Blocks()184 with app:185 gr.Markdown("# English & Chinese & Japanese Anime TTS\n\n"186 "\n\n"187 "Including Japanese TTS & Trilingual TTS, speakers are all anime characters. \n\n包含一个纯日语TTS和一个中日英三语TTS模型，主要为二次元角色。\n\n"188 "If you have any suggestions or bug reports, feel free to open discussion in [Community](https://huggingface.co/spaces/Plachta/VITS-Umamusume-voice-synthesizer/discussions).\n\n"189 "若有bug反馈或建议，请在[Community](https://huggingface.co/spaces/Plachta/VITS-Umamusume-voice-synthesizer/discussions)下开启一个新的Discussion。 \n\n"190 )191 with gr.Tabs():192 with gr.TabItem("TTS"):193 with gr.Tabs():194 for i, (name, description, speakers, lang, example, symbols, tts_fn, to_symbol_fn) in enumerate(195 models_tts):196 with gr.TabItem(name):197 gr.Markdown(description)198 with gr.Row():199 with gr.Column():200 textbox = gr.TextArea(label="Text",201 placeholder="Type your sentence here (Maximum 150 words)",202 value="こんにちわ。", elem_id=f"tts-input")203 with gr.Accordion(label="Phoneme Input", open=False):204 temp_text_var = gr.Variable()205 symbol_input = gr.Checkbox(value=False, label="Symbol input")206 symbol_list = gr.Dataset(label="Symbol list", components=[textbox],207 samples=[[x] for x in symbols],208 elem_id=f"symbol-list")209 symbol_list_json = gr.Json(value=symbols, visible=False)210 symbol_input.change(to_symbol_fn,211 [symbol_input, textbox, temp_text_var],212 [textbox, temp_text_var])213 symbol_list.click(None, [symbol_list, symbol_list_json], textbox,214 _js=f"""215 (i, symbols, text) => {{216 let root = document.querySelector("body > gradio-app");217 if (root.shadowRoot != null)218 root = root.shadowRoot;219 let text_input = root.querySelector("#tts-input").querySelector("textarea");220 let startPos = text_input.selectionStart;221 let endPos = text_input.selectionEnd;222 let oldTxt = text_input.value;223 let result = oldTxt.substring(0, startPos) + symbols[i] + oldTxt.substring(endPos);224 text_input.value = result;225 let x = window.scrollX, y = window.scrollY;226 text_input.focus();227 text_input.selectionStart = startPos + symbols[i].length;228 text_input.selectionEnd = startPos + symbols[i].length;229 text_input.blur();230 window.scrollTo(x, y);231 232 text = text_input.value;233 234 return text;235 }}""")236 # select character237 char_dropdown = gr.Dropdown(choices=speakers, value=speakers[0], label='character')238 language_dropdown = gr.Dropdown(choices=lang, value=lang[0], label='language')239 duration_slider = gr.Slider(minimum=0.1, maximum=5, value=1, step=0.1,240 label='速度 Speed')241 with gr.Column():242 text_output = gr.Textbox(label="Message")243 audio_output = gr.Audio(label="Output Audio", elem_id="tts-audio")244 btn = gr.Button("Generate!")245 btn.click(tts_fn,246 inputs=[textbox, char_dropdown, language_dropdown, duration_slider,247 symbol_input],248 outputs=[text_output, audio_output])249 gr.Examples(250 examples=example,251 inputs=[textbox, char_dropdown, language_dropdown,252 duration_slider, symbol_input],253 outputs=[text_output, audio_output],254 fn=tts_fn255 )256 app.queue(concurrency_count=3).launch(show_api=False, share=args.share)