import argparse import json import os import re import tempfile import logging logging.getLogger('numba').setLevel(logging.WARNING) import ONNXVITS_infer import librosa import numpy as np import torch from torch import no_grad, LongTensor import commons import utils import gradio as gr import gradio.utils as gr_utils import gradio.processing_utils as gr_processing_utils from models import SynthesizerTrn from text import text_to_sequence, _clean_text from text.symbols import symbols from mel_processing import spectrogram_torch import translators.server as tss import psutil from datetime import datetime from text.cleaners import japanese_cleaners from gradio import routes def audio_postprocess(self, y): if y is None: return None if gr_utils.validate_url(y): file = gr_processing_utils.download_to_file(y, dir=self.temp_dir) elif isinstance(y, tuple): sample_rate, data = y file = tempfile.NamedTemporaryFile( suffix=".wav", dir=self.temp_dir, delete=False ) gr_processing_utils.audio_to_file(sample_rate, data, file.name) else: file = gr_processing_utils.create_tmp_copy_of_file(y, dir=self.temp_dir) return gr_processing_utils.encode_url_or_file_to_base64(file.name) gr.Audio.postprocess = audio_postprocess limitation = os.getenv("SYSTEM") == "spaces" # limit text and audio length in huggingface spaces languages = ['日本語', '简体中文', 'English'] characters = ['0:特别周', '1:无声铃鹿', '2:东海帝王', '3:丸善斯基', '4:富士奇迹', '5:小栗帽', '6:黄金船', '7:伏特加', '8:大和赤骥', '9:大树快车', '10:草上飞', '11:菱亚马逊', '12:目白麦昆', '13:神鹰', '14:好歌剧', '15:成田白仁', '16:鲁道夫象征', '17:气槽', '18:爱丽数码', '19:青云天空', '20:玉藻十字', '21:美妙姿势', '22:琵琶晨光', '23:重炮', '24:曼城茶座', '25:美普波旁', '26:目白雷恩', '27:菱曙', '28:雪之美人', '29:米浴', '30:艾尼斯风神', '31:爱丽速子', '32:爱慕织姬', '33:稻荷一', '34:胜利奖券', '35:空中神宫', '36:荣进闪耀', '37:真机伶', '38:川上公主', '39:黄金城市', '40:樱花进王', '41:采珠', '42:新光风', '43:东商变革', '44:超级小溪', '45:醒目飞鹰', '46:荒漠英雄', '47:东瀛佐敦', '48:中山庆典', '49:成田大进', '50:西野花', '51:春乌拉拉', '52:青竹回忆', '53:微光飞驹', '54:美丽周日', '55:待兼福来', '56:Mr.C.B', '57:名将怒涛', '58:目白多伯', '59:优秀素质', '60:帝王光环', '61:待兼诗歌剧', '62:生野狄杜斯', '63:目白善信', '64:大拓太阳神', '65:双涡轮', '66:里见光钻', '67:北部玄驹', '68:樱花千代王', '69:天狼星象征', '70:目白阿尔丹', '71:八重无敌', '72:鹤丸刚志', '73:目白光明', '74:樱花桂冠', '75:成田路', '76:也文摄辉', '77:吉兆', '78:谷野美酒', '79:第一红宝石', '80:真弓快车', '81:骏川手纲', '82:凯斯奇迹', '83:小林历奇', '84:北港火山', '85:奇锐骏', '86:秋川理事长'] def show_memory_info(hint): pid = os.getpid() p = psutil.Process(pid) info = p.memory_info() memory = info.rss / 1024.0 / 1024 print("{} 内存占用: {} MB".format(hint, memory)) def text_to_phoneme(text, symbols, is_symbol): _symbol_to_id = {s: i for i, s in enumerate(symbols)} sequence = "" if not is_symbol: clean_text = japanese_cleaners(text) else: clean_text = text for symbol in clean_text: if symbol not in _symbol_to_id.keys(): continue symbol_id = _symbol_to_id[symbol] sequence += symbol return sequence def get_text(text, hps, is_symbol): text_norm = text_to_sequence(text, hps.symbols, [] if is_symbol else hps.data.text_cleaners) if hps.data.add_blank: text_norm = commons.intersperse(text_norm, 0) text_norm = LongTensor(text_norm) return text_norm hps = utils.get_hparams_from_file("./configs/uma87.json") symbols = hps.symbols net_g = ONNXVITS_infer.SynthesizerTrn( len(hps.symbols), hps.data.filter_length // 2 + 1, hps.train.segment_size // hps.data.hop_length, n_speakers=hps.data.n_speakers, **hps.model) _ = net_g.eval() _ = utils.load_checkpoint("pretrained_models/G_1153000.pth", net_g) def to_symbol_fn(is_symbol_input, input_text, temp_text): return (_clean_text(input_text, hps.data.text_cleaners), input_text) if is_symbol_input \ else (temp_text, temp_text) def infer(text_raw, character, language, duration, noise_scale, noise_scale_w, is_symbol): """ infer """ # check character & duraction parameter if language not in languages: print("Error: No such language\n") return "Error: No such language", None, None, None if character not in characters: print("Error: No such character\n") return "Error: No such character", None, None, None # check text length if limitation: text_len = len(text_raw) if is_symbol else len(re.sub("\[([A-Z]{2})\]", "", text_raw)) max_len = 150 if is_symbol: max_len *= 3 if text_len > max_len: print(f"Refused: Text too long ({text_len}).") return "Error: Text is too long", None, None, None if text_len == 0: print("Refused: Text length is zero.") return "Error: Please input text!", None, None, None if is_symbol: text = text_raw elif language == '日本語': text = text_raw elif language == '简体中文': text = tss.google(text_raw, from_language='zh', to_language='ja') elif language == 'English': text = tss.google(text_raw, from_language='en', to_language='ja') char_id = int(character.split(':')[0]) stn_tst = get_text(text, hps, is_symbol) with torch.no_grad(): x_tst = stn_tst.unsqueeze(0) x_tst_lengths = torch.LongTensor([stn_tst.size(0)]) sid = torch.LongTensor([char_id]) try: jp2phoneme = text_to_phoneme(text, hps.symbols, is_symbol) durations = net_g.predict_duration(x_tst, x_tst_lengths, sid=sid, noise_scale=noise_scale, noise_scale_w=noise_scale_w, length_scale=duration) char_dur_list = [] for i, char in enumerate(jp2phoneme): char_pos = i * 2 + 1 char_dur = durations[char_pos] char_dur_list.append(char_dur) except IndexError: print("Refused: Phoneme input contains non-phoneme character.") return "Error: You can only input phoneme under phoneme input model", None, None, None char_spacing_dur_list = [] char_spacings = [] for i in range(len(durations)): if i % 2 == 0: # spacing char_spacings.append("spacing") elif i % 2 == 1: # char char_spacings.append(jp2phoneme[int((i - 1) / 2)]) char_spacing_dur_list.append(int(durations[i])) # convert duration information to string duration_info_str = "" for i in range(len(char_spacings)): if i == len(char_spacings) - 1: duration_info_str += "(" + str(char_spacing_dur_list[i]) + ")" elif char_spacings[i] == "spacing": duration_info_str += "(" + str(char_spacing_dur_list[i]) + ")" + ", " else: duration_info_str += char_spacings[i] + ":" + str(char_spacing_dur_list[i]) audio = net_g.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=noise_scale, noise_scale_w=noise_scale_w, length_scale=duration)[0][0,0].data.float().numpy() currentDateAndTime = datetime.now() print(f"\nCharacter {character} inference successful: {text}") if language != '日本語': print(f"translate from {language}: {text_raw}") show_memory_info(str(currentDateAndTime) + " infer调用后") return (text,(22050, audio), jp2phoneme, duration_info_str) def infer_from_phoneme_dur(duration_info_str, character, duration, noise_scale, noise_scale_w): """ infer from phoneme dur """ try: phonemes = duration_info_str.split(", ") recons_durs = [] recons_phonemes = "" for i, item in enumerate(phonemes): if i == 0: recons_durs.append(int(item.strip("()"))) else: phoneme_n_dur, spacing_dur = item.split("(") recons_phonemes += phoneme_n_dur.split(":")[0] recons_durs.append(int(phoneme_n_dur.split(":")[1])) recons_durs.append(int(spacing_dur.strip(")"))) except ValueError: return ("Error: Format must not be changed!", None) except AssertionError: return ("Error: Format must not be changed!", None) char_id = int(character.split(':')[0]) stn_tst = get_text(recons_phonemes, hps, is_symbol=True) with torch.no_grad(): x_tst = stn_tst.unsqueeze(0) x_tst_lengths = torch.LongTensor([stn_tst.size(0)]) sid = torch.LongTensor([char_id]) audio = net_g.infer_with_duration(x_tst, x_tst_lengths, w_ceil=recons_durs, sid=sid, noise_scale=noise_scale, noise_scale_w=noise_scale_w, length_scale=duration)[0][0, 0].data.cpu().float().numpy() print(f"\nCharacter {character} inference successful: {recons_phonemes}, from {duration_info_str}") return (recons_phonemes, (22050, audio)) download_audio_js = """ () =>{{ let root = document.querySelector("body > gradio-app"); if (root.shadowRoot != null) root = root.shadowRoot; let audio = root.querySelector("#{audio_id}").querySelector("audio"); if (audio == undefined) return; audio = audio.src; let oA = document.createElement("a"); oA.download = Math.floor(Math.random()*100000000)+'.wav'; oA.href = audio; document.body.appendChild(oA); oA.click(); oA.remove(); }} """ def monkey_patch(): def postprocess(self, y): """ Any postprocessing needed to be performed on a block context. """ return y gr.blocks.BlockContext.postprocess = postprocess def get_types(cls_set: List[Type], component: str): docset = [] types = [] if component == "input": for cls in cls_set: doc = inspect.getdoc(cls) doc_lines = doc.split("\n") docset.append(doc_lines[1].split(":")[-1]) types.append(doc_lines[1].split(")")[0].split("(")[-1]) else: for cls in cls_set: doc = inspect.getdoc(cls) doc_lines = doc.split("\n") docset.append(doc_lines[-1].split(":")[-1]) types.append(doc_lines[-1].split(")")[0].split("(")[-1]) return docset, types routes.get_types = get_types if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--share", action="store_true", default=False, help="share gradio app") args = parser.parse_args() app = gr.Blocks() with app: gr.Markdown("# Umamusume voice synthesizer 赛马娘语音合成器\n\n" "![visitor badge](https://visitor-badge.glitch.me/badge?page_id=Plachta.VITS-Umamusume-voice-synthesizer)\n\n" "This synthesizer is created based on [VITS](https://arxiv.org/abs/2106.06103) model, trained on voice data extracted from mobile game Umamusume Pretty Derby \n\n" "这个合成器是基于VITS文本到语音模型,在从手游《賽馬娘:Pretty Derby》解包的语音数据上训练得到。[Dataset Link](https://huggingface.co/datasets/Plachta/Umamusume-voice-text-pairs/tree/main)\n\n" "[introduction video / 模型介绍视频](https://www.bilibili.com/video/BV1T84y1e7p5/?vd_source=6d5c00c796eff1cbbe25f1ae722c2f9f#reply607277701)\n\n" "You may duplicate this space or [open in Colab](https://colab.research.google.com/drive/1J2Vm5dczTF99ckyNLXV0K-hQTxLwEaj5?usp=sharing) to run it privately and without any queue.\n\n" "您可以复制该空间至私人空间运行或打开[Google Colab](https://colab.research.google.com/drive/1J2Vm5dczTF99ckyNLXV0K-hQTxLwEaj5?usp=sharing)在线运行。\n\n" "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" "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" "If your input language is not Japanese, it will be translated to Japanese by Google translator, but accuracy is not guaranteed.\n\n" "如果您的输入语言不是日语,则会由谷歌翻译自动翻译为日语,但是准确性不能保证。\n\n" ) with gr.Row(): with gr.Column(): # We instantiate the Textbox class textbox = gr.TextArea(label="Text", placeholder="Type your sentence here (Maximum 150 words)", value="こんにちわ。", elem_id=f"tts-input") with gr.Accordion(label="Phoneme Input", open=False): temp_text_var = gr.Variable() symbol_input = gr.Checkbox(value=False, label="Symbol input") symbol_list = gr.Dataset(label="Symbol list", components=[textbox], samples=[[x] for x in symbols], elem_id=f"symbol-list") symbol_list_json = gr.Json(value=symbols, visible=False) symbol_input.change(to_symbol_fn, [symbol_input, textbox, temp_text_var], [textbox, temp_text_var]) symbol_list.click(None, [symbol_list, symbol_list_json], textbox, _js=f""" (i, symbols, text) => {{ let root = document.querySelector("body > gradio-app"); if (root.shadowRoot != null) root = root.shadowRoot; let text_input = root.querySelector("#tts-input").querySelector("textarea"); let startPos = text_input.selectionStart; let endPos = text_input.selectionEnd; let oldTxt = text_input.value; let result = oldTxt.substring(0, startPos) + symbols[i] + oldTxt.substring(endPos); text_input.value = result; let x = window.scrollX, y = window.scrollY; text_input.focus(); text_input.selectionStart = startPos + symbols[i].length; text_input.selectionEnd = startPos + symbols[i].length; text_input.blur(); window.scrollTo(x, y); text = text_input.value; return text; }}""") # select character char_dropdown = gr.Dropdown(choices=characters, value = "0:特别周", label='character') language_dropdown = gr.Dropdown(choices=languages, value = "日本語", label='language') duration_slider = gr.Slider(minimum=0.1, maximum=5, value=1, step=0.1, label='时长 Duration') noise_scale_slider = gr.Slider(minimum=0.1, maximum=5, value=0.667, step=0.001, label='噪声比例 noise_scale') noise_scale_w_slider = gr.Slider(minimum=0.1, maximum=5, value=0.8, step=0.1, label='噪声偏差 noise_scale_w') with gr.Column(): text_output = gr.Textbox(label="Output Text") phoneme_output = gr.Textbox(label="Output Phonemes", interactive=False) audio_output = gr.Audio(label="Output Audio", elem_id="tts-audio") btn = gr.Button("Generate!") cus_dur_gn_btn = gr.Button("Regenerate with custom phoneme durations") download = gr.Button("Download Audio") download.click(None, [], [], _js=download_audio_js.format(audio_id="tts-audio"), api_name="download_audio") with gr.Accordion(label="Speaking Pace Control", open=True): duration_output = gr.Textbox(label="Duration of each phoneme", placeholder="After you generate a sentence, the detailed information of each phoneme's duration will be presented here.", interactive = True) gr.Markdown( "The number after the : mark represents the length of each phoneme in the generated audio, while the number inside ( ) represents the lenght of spacing between each phoneme and its next phoneme. " "You can manually change the numbers to adjust the length of each phoneme, so that speaking pace can be completely controlled. " "Note that these numbers should be integers only. \n\n(1 represents a length of 0.01161 seconds)\n\n" "音素冒号后的数字代表音素在生成音频中的长度,( )内的数字代表每个音素与下一个音素之间间隔的长度。" "您可以手动修改这些数字来控制每个音素以及间隔的长度,从而完全控制合成音频的说话节奏。" "注意这些数字只能是整数。 \n\n(1 代表 0.01161 秒的长度)\n\n" ) monkey_patch() btn.click(infer, inputs=[textbox, char_dropdown, language_dropdown, duration_slider, noise_scale_slider, noise_scale_w_slider, symbol_input], outputs=[text_output, audio_output, phoneme_output, duration_output])#, api_name="1") cus_dur_gn_btn.click(infer_from_phoneme_dur, inputs=[duration_output, char_dropdown, duration_slider, noise_scale_slider, noise_scale_w_slider], outputs=[phoneme_output, audio_output])#, api_name="2") examples = [['haa\u2193......haa\u2193......haa\u2193......haa\u2193......haa\u2193......haa\u2193......haa\u2193......haa\u2193......haa\u2193......haa\u2193......haa\u2193......haa\u2193......', '29:米浴', '日本語', 1, 0.667, 0.8, True], ['お疲れ様です,トレーナーさん。', '1:无声铃鹿', '日本語', 1, 0.667, 0.8, False], ['張り切っていこう!', '67:北部玄驹', '日本語', 1, 0.667, 0.8, False], ['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。', '10:草上飞', '日本語', 1, 0.667, 0.8, False], ['授業中に出しだら,学校生活終わるですわ。', '12:目白麦昆', '日本語', 1, 0.667, 0.8, False], ['お帰りなさい,お兄様!', '29:米浴', '日本語', 1, 0.667, 0.8, False], ['私の処女をもらっでください!', '29:米浴', '日本語', 1, 0.667, 0.8, False]] gr.Examples( examples=examples, inputs=[textbox, char_dropdown, language_dropdown, duration_slider, noise_scale_slider,noise_scale_w_slider, symbol_input], outputs=[text_output, audio_output], fn=infer ) gr.Markdown("# Updates Logs 更新日志:\n\n" "2023/1/24:\n\n" "Improved the format of phoneme length control.\n\n" "改善了音素控制的格式。\n\n" "2023/1/24:\n\n" "Added more precise control on pace of speaking by modifying the duration of each phoneme.\n\n" "增加了对说话节奏的音素级控制。\n\n" "2023/1/13:\n\n" "Added one example of phoneme input.\n\n" "增加了音素输入的example(米浴喘气)\n\n" "2023/1/12:\n\n" "Added phoneme input, which enables more precise control on output audio.\n\n" "增加了音素输入的功能,可以对语气和语调做到一定程度的精细控制。\n\n" "Adjusted UI arrangements.\n\n" "调整了UI的布局。\n\n" "2023/1/10:\n\n" "Dataset used for training is now uploaded to [here](https://huggingface.co/datasets/Plachta/Umamusume-voice-text-pairs/tree/main)\n\n" "数据集已上传,您可以在[这里](https://huggingface.co/datasets/Plachta/Umamusume-voice-text-pairs/tree/main)下载。\n\n" "2023/1/9:\n\n" "Model inference has been fully converted to onnxruntime. There will be no more Runtime Error: Memory Limit Exceeded\n\n" "模型推理已全面转为onnxruntime,现在不会出现Runtime Error: Memory Limit Exceeded了。\n\n" "Now integrated to [Moe-tts](https://huggingface.co/spaces/skytnt/moe-tts) collection.\n\n" "现已加入[Moe-tts](https://huggingface.co/spaces/skytnt/moe-tts)模型大全。\n\n" ) ifa = gr.Interface(lambda: None, inputs=[textbox], outputs=[text_output]) app.queue(concurrency_count=3).launch(show_api=True, share=args.share)