diff --git a/GPT_SoVITS/inference_gui.py b/GPT_SoVITS/inference_gui.py index fd2dae86..d5238049 100644 --- a/GPT_SoVITS/inference_gui.py +++ b/GPT_SoVITS/inference_gui.py @@ -1,480 +1,13 @@ -import os,re,logging -logging.getLogger("markdown_it").setLevel(logging.ERROR) -logging.getLogger("urllib3").setLevel(logging.ERROR) -logging.getLogger("httpcore").setLevel(logging.ERROR) -logging.getLogger("httpx").setLevel(logging.ERROR) -logging.getLogger("asyncio").setLevel(logging.ERROR) - -logging.getLogger("charset_normalizer").setLevel(logging.ERROR) -logging.getLogger("torchaudio._extension").setLevel(logging.ERROR) -import pdb - -if os.path.exists("./gweight.txt"): - with open("./gweight.txt", 'r',encoding="utf-8") as file: - gweight_data = file.read() - gpt_path = os.environ.get( - "gpt_path", gweight_data) -else: - gpt_path = os.environ.get( - "gpt_path", "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt") - -if os.path.exists("./sweight.txt"): - with open("./sweight.txt", 'r',encoding="utf-8") as file: - sweight_data = file.read() - sovits_path = os.environ.get("sovits_path", sweight_data) -else: - sovits_path = os.environ.get("sovits_path", "GPT_SoVITS/pretrained_models/s2G488k.pth") -# gpt_path = os.environ.get( -# "gpt_path", "pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt" -# ) -# sovits_path = os.environ.get("sovits_path", "pretrained_models/s2G488k.pth") -cnhubert_base_path = os.environ.get( - "cnhubert_base_path", "GPT_SoVITS/pretrained_models/chinese-hubert-base" -) -bert_path = os.environ.get( - "bert_path", "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large" -) -infer_ttswebui = os.environ.get("infer_ttswebui", 9872) -infer_ttswebui = int(infer_ttswebui) -is_share = os.environ.get("is_share", "False") -is_share=eval(is_share) -if "_CUDA_VISIBLE_DEVICES" in os.environ: - os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["_CUDA_VISIBLE_DEVICES"] -is_half = eval(os.environ.get("is_half", "True")) -import gradio as gr -from transformers import AutoModelForMaskedLM, AutoTokenizer -import numpy as np -import librosa,torch -from feature_extractor import cnhubert -cnhubert.cnhubert_base_path=cnhubert_base_path - import sys from PyQt5.QtCore import QEvent from PyQt5.QtWidgets import QApplication, QMainWindow, QLabel, QLineEdit, QPushButton, QTextEdit from PyQt5.QtWidgets import QGridLayout, QVBoxLayout, QWidget, QFileDialog, QStatusBar, QComboBox import soundfile as sf -from module.models import SynthesizerTrn -from AR.models.t2s_lightning_module import Text2SemanticLightningModule -from text import cleaned_text_to_sequence -from text.cleaner import clean_text -from time import time as ttime -from module.mel_processing import spectrogram_torch -from my_utils import load_audio from tools.i18n.i18n import I18nAuto i18n = I18nAuto() -os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 确保直接启动推理UI时也能够设置。 - -if torch.cuda.is_available(): - device = "cuda" -elif torch.backends.mps.is_available(): - device = "mps" -else: - device = "cpu" - -tokenizer = AutoTokenizer.from_pretrained(bert_path) -bert_model = AutoModelForMaskedLM.from_pretrained(bert_path) -if is_half == True: - bert_model = bert_model.half().to(device) -else: - bert_model = bert_model.to(device) - -def get_bert_feature(text, word2ph): - with torch.no_grad(): - inputs = tokenizer(text, return_tensors="pt") - for i in inputs: - inputs[i] = inputs[i].to(device) - res = bert_model(**inputs, output_hidden_states=True) - res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()[1:-1] - assert len(word2ph) == len(text) - phone_level_feature = [] - for i in range(len(word2ph)): - repeat_feature = res[i].repeat(word2ph[i], 1) - phone_level_feature.append(repeat_feature) - phone_level_feature = torch.cat(phone_level_feature, dim=0) - return phone_level_feature.T - -class DictToAttrRecursive(dict): - def __init__(self, input_dict): - super().__init__(input_dict) - for key, value in input_dict.items(): - if isinstance(value, dict): - value = DictToAttrRecursive(value) - self[key] = value - setattr(self, key, value) - - def __getattr__(self, item): - try: - return self[item] - except KeyError: - raise AttributeError(f"Attribute {item} not found") - - def __setattr__(self, key, value): - if isinstance(value, dict): - value = DictToAttrRecursive(value) - super(DictToAttrRecursive, self).__setitem__(key, value) - super().__setattr__(key, value) - - def __delattr__(self, item): - try: - del self[item] - except KeyError: - raise AttributeError(f"Attribute {item} not found") - - -ssl_model = cnhubert.get_model() -if is_half == True: - ssl_model = ssl_model.half().to(device) -else: - ssl_model = ssl_model.to(device) - -def change_sovits_weights(sovits_path): - global vq_model,hps - dict_s2=torch.load(sovits_path,map_location="cpu") - hps=dict_s2["config"] - hps = DictToAttrRecursive(hps) - hps.model.semantic_frame_rate = "25hz" - vq_model = SynthesizerTrn( - hps.data.filter_length // 2 + 1, - hps.train.segment_size // hps.data.hop_length, - n_speakers=hps.data.n_speakers, - **hps.model - ) - if("pretrained"not in sovits_path): - del vq_model.enc_q - if is_half == True: - vq_model = vq_model.half().to(device) - else: - vq_model = vq_model.to(device) - vq_model.eval() - print(vq_model.load_state_dict(dict_s2["weight"], strict=False)) - with open("./sweight.txt","w",encoding="utf-8")as f:f.write(sovits_path) -change_sovits_weights(sovits_path) - -def change_gpt_weights(gpt_path): - global hz,max_sec,t2s_model,config - hz = 50 - dict_s1 = torch.load(gpt_path, map_location="cpu") - config = dict_s1["config"] - max_sec = config["data"]["max_sec"] - t2s_model = Text2SemanticLightningModule(config, "****", is_train=False) - t2s_model.load_state_dict(dict_s1["weight"]) - if is_half == True: - t2s_model = t2s_model.half() - t2s_model = t2s_model.to(device) - t2s_model.eval() - total = sum([param.nelement() for param in t2s_model.parameters()]) - print("Number of parameter: %.2fM" % (total / 1e6)) - with open("./gweight.txt","w",encoding="utf-8")as f:f.write(gpt_path) -change_gpt_weights(gpt_path) - -def get_spepc(hps, filename): - audio = load_audio(filename, int(hps.data.sampling_rate)) - audio = torch.FloatTensor(audio) - audio_norm = audio - audio_norm = audio_norm.unsqueeze(0) - spec = spectrogram_torch( - audio_norm, - hps.data.filter_length, - hps.data.sampling_rate, - hps.data.hop_length, - hps.data.win_length, - center=False, - ) - return spec - - -dict_language={ - i18n("中文"):"zh", - i18n("英文"):"en", - i18n("日文"):"ja" -} - - -def splite_en_inf(sentence, language): - pattern = re.compile(r'[a-zA-Z. ]+') - textlist = [] - langlist = [] - pos = 0 - for match in pattern.finditer(sentence): - start, end = match.span() - if start > pos: - textlist.append(sentence[pos:start]) - langlist.append(language) - textlist.append(sentence[start:end]) - langlist.append("en") - pos = end - if pos < len(sentence): - textlist.append(sentence[pos:]) - langlist.append(language) - - return textlist, langlist - - -def clean_text_inf(text, language): - phones, word2ph, norm_text = clean_text(text, language) - phones = cleaned_text_to_sequence(phones) - - return phones, word2ph, norm_text - - -def get_bert_inf(phones, word2ph, norm_text, language): - if language == "zh": - bert = get_bert_feature(norm_text, word2ph).to(device) - else: - bert = torch.zeros( - (1024, len(phones)), - dtype=torch.float16 if is_half == True else torch.float32, - ).to(device) - - return bert - - -def nonen_clean_text_inf(text, language): - textlist, langlist = splite_en_inf(text, language) - phones_list = [] - word2ph_list = [] - norm_text_list = [] - for i in range(len(textlist)): - lang = langlist[i] - phones, word2ph, norm_text = clean_text_inf(textlist[i], lang) - phones_list.append(phones) - if lang == "en" or "ja": - pass - else: - word2ph_list.append(word2ph) - norm_text_list.append(norm_text) - print(word2ph_list) - phones = sum(phones_list, []) - word2ph = sum(word2ph_list, []) - norm_text = ' '.join(norm_text_list) - - return phones, word2ph, norm_text - - -def nonen_get_bert_inf(text, language): - textlist, langlist = splite_en_inf(text, language) - print(textlist) - print(langlist) - bert_list = [] - for i in range(len(textlist)): - text = textlist[i] - lang = langlist[i] - phones, word2ph, norm_text = clean_text_inf(text, lang) - bert = get_bert_inf(phones, word2ph, norm_text, lang) - bert_list.append(bert) - bert = torch.cat(bert_list, dim=1) - - return bert - -splits = {"，","。","？","！",",",".","?","!","~",":","：","—","…",} -def get_first(text): - pattern = "[" + "".join(re.escape(sep) for sep in splits) + "]" - text = re.split(pattern, text)[0].strip() - return text - -def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language,how_to_cut=i18n("不切")): - t0 = ttime() - prompt_text = prompt_text.strip("\n") - if(prompt_text[-1]not in splits):prompt_text+="。"if prompt_text!="en"else "." - text = text.strip("\n") - if(len(get_first(text))<4):text+="。"if text!="en"else "." - zero_wav = np.zeros( - int(hps.data.sampling_rate * 0.3), - dtype=np.float16 if is_half == True else np.float32, - ) - with torch.no_grad(): - wav16k, sr = librosa.load(ref_wav_path, sr=16000) - if(wav16k.shape[0]>160000 or wav16k.shape[0]<48000): - raise OSError(i18n("参考音频在3~10秒范围外，请更换！")) - wav16k = torch.from_numpy(wav16k) - zero_wav_torch = torch.from_numpy(zero_wav) - if is_half == True: - wav16k = wav16k.half().to(device) - zero_wav_torch = zero_wav_torch.half().to(device) - else: - wav16k = wav16k.to(device) - zero_wav_torch = zero_wav_torch.to(device) - wav16k=torch.cat([wav16k,zero_wav_torch]) - ssl_content = ssl_model.model(wav16k.unsqueeze(0))[ - "last_hidden_state" - ].transpose( - 1, 2 - ) # .float() - codes = vq_model.extract_latent(ssl_content) - prompt_semantic = codes[0, 0] - t1 = ttime() - prompt_language = dict_language[prompt_language] - text_language = dict_language[text_language] - - if prompt_language == "en": - phones1, word2ph1, norm_text1 = clean_text_inf(prompt_text, prompt_language) - else: - phones1, word2ph1, norm_text1 = nonen_clean_text_inf(prompt_text, prompt_language) - if(how_to_cut==i18n("凑四句一切")):text=cut1(text) - elif(how_to_cut==i18n("凑50字一切")):text=cut2(text) - elif(how_to_cut==i18n("按中文句号。切")):text=cut3(text) - elif(how_to_cut==i18n("按英文句号.切")):text=cut4(text) - text = text.replace("\n\n","\n").replace("\n\n","\n").replace("\n\n","\n") - if(text[-1]not in splits):text+="。"if text_language!="en"else "." - texts=text.split("\n") - audio_opt = [] - if prompt_language == "en": - bert1 = get_bert_inf(phones1, word2ph1, norm_text1, prompt_language) - else: - bert1 = nonen_get_bert_inf(prompt_text, prompt_language) - - for text in texts: - # 解决输入目标文本的空行导致报错的问题 - if (len(text.strip()) == 0): - continue - if text_language == "en": - phones2, word2ph2, norm_text2 = clean_text_inf(text, text_language) - else: - phones2, word2ph2, norm_text2 = nonen_clean_text_inf(text, text_language) - - if text_language == "en": - bert2 = get_bert_inf(phones2, word2ph2, norm_text2, text_language) - else: - bert2 = nonen_get_bert_inf(text, text_language) - - bert = torch.cat([bert1, bert2], 1) - - all_phoneme_ids = torch.LongTensor(phones1 + phones2).to(device).unsqueeze(0) - bert = bert.to(device).unsqueeze(0) - all_phoneme_len = torch.tensor([all_phoneme_ids.shape[-1]]).to(device) - prompt = prompt_semantic.unsqueeze(0).to(device) - t2 = ttime() - with torch.no_grad(): - # pred_semantic = t2s_model.model.infer( - pred_semantic, idx = t2s_model.model.infer_panel( - all_phoneme_ids, - all_phoneme_len, - prompt, - bert, - # prompt_phone_len=ph_offset, - top_k=config["inference"]["top_k"], - early_stop_num=hz * max_sec, - ) - t3 = ttime() - # print(pred_semantic.shape,idx) - pred_semantic = pred_semantic[:, -idx:].unsqueeze( - 0 - ) # .unsqueeze(0)#mq要多unsqueeze一次 - refer = get_spepc(hps, ref_wav_path) # .to(device) - if is_half == True: - refer = refer.half().to(device) - else: - refer = refer.to(device) - # audio = vq_model.decode(pred_semantic, all_phoneme_ids, refer).detach().cpu().numpy()[0, 0] - audio = ( - vq_model.decode( - pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refer - ) - .detach() - .cpu() - .numpy()[0, 0] - ) ###试试重建不带上prompt部分 - audio_opt.append(audio) - audio_opt.append(zero_wav) - t4 = ttime() - print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3)) - yield hps.data.sampling_rate, (np.concatenate(audio_opt, 0) * 32768).astype( - np.int16 - ) - -def split(todo_text): - todo_text = todo_text.replace("……", "。").replace("——", "，") - if todo_text[-1] not in splits: - todo_text += "。" - i_split_head = i_split_tail = 0 - len_text = len(todo_text) - todo_texts = [] - while 1: - if i_split_head >= len_text: - break # 结尾一定有标点，所以直接跳出即可，最后一段在上次已加入 - if todo_text[i_split_head] in splits: - i_split_head += 1 - todo_texts.append(todo_text[i_split_tail:i_split_head]) - i_split_tail = i_split_head - else: - i_split_head += 1 - return todo_texts - - -def cut1(inp): - inp = inp.strip("\n") - inps = split(inp) - split_idx = list(range(0, len(inps), 4)) - split_idx[-1] = None - if len(split_idx) > 1: - opts = [] - for idx in range(len(split_idx) - 1): - opts.append("".join(inps[split_idx[idx] : split_idx[idx + 1]])) - else: - opts = [inp] - return "\n".join(opts) - - -def cut2(inp): - inp = inp.strip("\n") - inps = split(inp) - if len(inps) < 2: - return inp - opts = [] - summ = 0 - tmp_str = "" - for i in range(len(inps)): - summ += len(inps[i]) - tmp_str += inps[i] - if summ > 50: - summ = 0 - opts.append(tmp_str) - tmp_str = "" - if tmp_str != "": - opts.append(tmp_str) - # print(opts) - if len(opts)>1 and len(opts[-1]) < 50: ##如果最后一个太短了，和前一个合一起 - opts[-2] = opts[-2] + opts[-1] - opts = opts[:-1] - return "\n".join(opts) - - -def cut3(inp): - inp = inp.strip("\n") - return "\n".join(["%s。" % item for item in inp.strip("。").split("。")]) -def cut4(inp): - inp = inp.strip("\n") - return "\n".join(["%s." % item for item in inp.strip(".").split(".")]) - -def custom_sort_key(s): - # 使用正则表达式提取字符串中的数字部分和非数字部分 - parts = re.split('(\d+)', s) - # 将数字部分转换为整数，非数字部分保持不变 - parts = [int(part) if part.isdigit() else part for part in parts] - return parts - -def change_choices(): - SoVITS_names, GPT_names = get_weights_names() - return {"choices": sorted(SoVITS_names,key=custom_sort_key), "__type__": "update"}, {"choices": sorted(GPT_names,key=custom_sort_key), "__type__": "update"} - -pretrained_sovits_name="GPT_SoVITS/pretrained_models/s2G488k.pth" -pretrained_gpt_name="GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt" -SoVITS_weight_root="SoVITS_weights" -GPT_weight_root="GPT_weights" -os.makedirs(SoVITS_weight_root,exist_ok=True) -os.makedirs(GPT_weight_root,exist_ok=True) - -def get_weights_names(): - SoVITS_names = [pretrained_sovits_name] - for name in os.listdir(SoVITS_weight_root): - if name.endswith(".pth"):SoVITS_names.append("%s/%s"%(SoVITS_weight_root,name)) - GPT_names = [pretrained_gpt_name] - for name in os.listdir(GPT_weight_root): - if name.endswith(".ckpt"): GPT_names.append("%s/%s"%(GPT_weight_root,name)) - return SoVITS_names,GPT_names -SoVITS_names,GPT_names = get_weights_names() +from GPT_SoVITS.inference_webui import change_gpt_weights, change_sovits_weights, get_tts_wav class GPTSoVITSGUI(QMainWindow): diff --git a/GPT_SoVITS/inference_webui.py b/GPT_SoVITS/inference_webui.py index f733fd0f..1868a122 100644 --- a/GPT_SoVITS/inference_webui.py +++ b/GPT_SoVITS/inference_webui.py @@ -1,11 +1,18 @@ +''' +按中英混合识别 +按日英混合识别 +多语种启动切分识别语种 +全部按中文识别 +全部按英文识别 +全部按日文识别 +''' import os, re, logging - +import LangSegment logging.getLogger("markdown_it").setLevel(logging.ERROR) logging.getLogger("urllib3").setLevel(logging.ERROR) logging.getLogger("httpcore").setLevel(logging.ERROR) logging.getLogger("httpx").setLevel(logging.ERROR) logging.getLogger("asyncio").setLevel(logging.ERROR) - logging.getLogger("charset_normalizer").setLevel(logging.ERROR) logging.getLogger("torchaudio._extension").setLevel(logging.ERROR) import pdb @@ -193,9 +200,12 @@ def get_spepc(hps, filename): dict_language = { - i18n("中文"): "zh", - i18n("英文"): "en", - i18n("日文"): "ja" + i18n("中文"): "all_zh",#全部按中文识别 + i18n("英文"): "en",#全部按英文识别#######不变 + i18n("日文"): "all_ja",#全部按日文识别 + i18n("中英混合"): "zh",#按中英混合识别####不变 + i18n("日英混合"): "ja",#按日英混合识别####不变 + i18n("多语种混合"): "auto",#多语种启动切分识别语种 } @@ -235,15 +245,15 @@ def splite_en_inf(sentence, language): def clean_text_inf(text, language): - phones, word2ph, norm_text = clean_text(text, language) + phones, word2ph, norm_text = clean_text(text, language.replace("all_","")) phones = cleaned_text_to_sequence(phones) - return phones, word2ph, norm_text - +dtype=torch.float16 if is_half == True else torch.float32 def get_bert_inf(phones, word2ph, norm_text, language): + language=language.replace("all_","") if language == "zh": - bert = get_bert_feature(norm_text, word2ph).to(device) + bert = get_bert_feature(norm_text, word2ph).to(device)#.to(dtype) else: bert = torch.zeros( (1024, len(phones)), @@ -254,7 +264,16 @@ def get_bert_inf(phones, word2ph, norm_text, language): def nonen_clean_text_inf(text, language): - textlist, langlist = splite_en_inf(text, language) + if(language!="auto"): + textlist, langlist = splite_en_inf(text, language) + else: + textlist=[] + langlist=[] + for tmp in LangSegment.getTexts(text): + langlist.append(tmp["lang"]) + textlist.append(tmp["text"]) + print(textlist) + print(langlist) phones_list = [] word2ph_list = [] norm_text_list = [] @@ -262,9 +281,7 @@ def nonen_clean_text_inf(text, language): lang = langlist[i] phones, word2ph, norm_text = clean_text_inf(textlist[i], lang) phones_list.append(phones) - if lang == "en" or "ja": - pass - else: + if lang == "zh": word2ph_list.append(word2ph) norm_text_list.append(norm_text) print(word2ph_list) @@ -276,7 +293,14 @@ def nonen_clean_text_inf(text, language): def nonen_get_bert_inf(text, language): - textlist, langlist = splite_en_inf(text, language) + if(language!="auto"): + textlist, langlist = splite_en_inf(text, language) + else: + textlist=[] + langlist=[] + for tmp in LangSegment.getTexts(text): + langlist.append(tmp["lang"]) + textlist.append(tmp["text"]) print(textlist) print(langlist) bert_list = [] @@ -300,6 +324,24 @@ def get_first(text): return text +def get_cleaned_text_fianl(text,language): + if language in {"en","all_zh","all_ja"}: + phones, word2ph, norm_text = clean_text_inf(text, language) + elif language in {"zh", "ja","auto"}: + phones, word2ph, norm_text = nonen_clean_text_inf(text, language) + return phones, word2ph, norm_text + +def get_bert_final(phones, word2ph, norm_text,language,device): + if text_language == "en": + bert = get_bert_inf(phones, word2ph, norm_text, text_language) + elif text_language in {"zh", "ja","auto"}: + bert = nonen_get_bert_inf(text, text_language) + elif text_language == "all_zh": + bert = get_bert_feature(norm_text, word2ph).to(device) + else: + bert = torch.zeros((1024, len(phones))).to(device) + return bert + def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切")): t0 = ttime() prompt_text = prompt_text.strip("\n") @@ -335,10 +377,9 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, t1 = ttime() prompt_language = dict_language[prompt_language] text_language = dict_language[text_language] - if prompt_language == "en": - phones1, word2ph1, norm_text1 = clean_text_inf(prompt_text, prompt_language) - else: - phones1, word2ph1, norm_text1 = nonen_clean_text_inf(prompt_text, prompt_language) + + phones1, word2ph1, norm_text1=get_cleaned_text_fianl(prompt_text, prompt_language) + if (how_to_cut == i18n("凑四句一切")): text = cut1(text) elif (how_to_cut == i18n("凑50字一切")): @@ -353,25 +394,16 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, print(i18n("实际输入的目标文本(切句后):"), text) texts = text.split("\n") audio_opt = [] - if prompt_language == "en": - bert1 = get_bert_inf(phones1, word2ph1, norm_text1, prompt_language) - else: - bert1 = nonen_get_bert_inf(prompt_text, prompt_language) + bert1=get_bert_final(phones1, word2ph1, norm_text1,prompt_language,device).to(dtype) + for text in texts: # 解决输入目标文本的空行导致报错的问题 if (len(text.strip()) == 0): continue if (text[-1] not in splits): text += "。" if text_language != "en" else "." print(i18n("实际输入的目标文本(每句):"), text) - if text_language == "en": - phones2, word2ph2, norm_text2 = clean_text_inf(text, text_language) - else: - phones2, word2ph2, norm_text2 = nonen_clean_text_inf(text, text_language) - - if text_language == "en": - bert2 = get_bert_inf(phones2, word2ph2, norm_text2, text_language) - else: - bert2 = nonen_get_bert_inf(text, text_language) + phones2, word2ph2, norm_text2 = get_cleaned_text_fianl(text, text_language) + bert2 = get_bert_final(phones2, word2ph2, norm_text2, text_language, device).to(dtype) bert = torch.cat([bert1, bert2], 1) @@ -557,7 +589,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app: with gr.Row(): text = gr.Textbox(label=i18n("需要合成的文本"), value="") text_language = gr.Dropdown( - label=i18n("需要合成的语种"), choices=[i18n("中文"), i18n("英文"), i18n("日文")], value=i18n("中文") + label=i18n("需要合成的语种"), choices=[i18n("中文"), i18n("英文"), i18n("日文"), i18n("中英混合"), i18n("日英混合"), i18n("多语种混合")], value=i18n("中文") ) how_to_cut = gr.Radio( label=i18n("怎么切"), diff --git a/GPT_SoVITS/text/cleaner.py b/GPT_SoVITS/text/cleaner.py index 8142f47d..92a18ebd 100644 --- a/GPT_SoVITS/text/cleaner.py +++ b/GPT_SoVITS/text/cleaner.py @@ -10,6 +10,9 @@ special = [ def clean_text(text, language): + if(language not in language_module_map): + language="en" + text=" " for special_s, special_l, target_symbol in special: if special_s in text and language == special_l: return clean_special(text, language, special_s, target_symbol) diff --git a/GPT_SoVITS/text/japanese.py b/GPT_SoVITS/text/japanese.py index 68112b96..a571467c 100644 --- a/GPT_SoVITS/text/japanese.py +++ b/GPT_SoVITS/text/japanese.py @@ -97,7 +97,7 @@ def text_normalize(text): return text # Copied from espnet https://github.com/espnet/espnet/blob/master/espnet2/text/phoneme_tokenizer.py -def pyopenjtalk_g2p_prosody(text: str, drop_unvoiced_vowels: bool = True) -> list[str]: +def pyopenjtalk_g2p_prosody(text, drop_unvoiced_vowels=True): """Extract phoneme + prosoody symbol sequence from input full-context labels. The algorithm is based on `Prosodic features control by symbols as input of diff --git a/docs/cn/Changelog_CN.md b/docs/cn/Changelog_CN.md index 8125ba61..8d56db0b 100644 --- a/docs/cn/Changelog_CN.md +++ b/docs/cn/Changelog_CN.md @@ -75,6 +75,12 @@ 3-增加按标点符号切分 +### 20240201更新 + +1-修复uvr5读取格式错误导致分离失败的问题 + +2-支持中日英混合多种文本自动切分识别语种 + todolist： @@ -84,4 +90,3 @@ todolist： 3-%百分号在文本里会导致error不能推理 还有 元/吨 会读成 元吨 而不是元每吨，像这类问题，是什么库来弄文本解析到语音的，应该怎么改善这个问题呀 -4-中日英、中英、日英 五种目标语言 diff --git a/tools/damo_asr/cmd-asr.py b/tools/damo_asr/cmd-asr.py index 70dd4d8f..9a107972 100644 --- a/tools/damo_asr/cmd-asr.py +++ b/tools/damo_asr/cmd-asr.py @@ -5,6 +5,7 @@ import sys,os,traceback from funasr import AutoModel dir=sys.argv[1] +if(dir[-1]=="/"):dir=dir[:-1] # opt_name=dir.split("\\")[-1].split("/")[-1] opt_name=os.path.basename(dir) diff --git a/tools/webui.py b/tools/webui.py new file mode 100644 index 00000000..41ec5887 --- /dev/null +++ b/tools/webui.py @@ -0,0 +1,178 @@ +import os +import traceback,gradio as gr +import logging +from tools.i18n.i18n import I18nAuto +i18n = I18nAuto() + +logger = logging.getLogger(__name__) +import librosa,ffmpeg +import soundfile as sf +import torch +import sys +from mdxnet import MDXNetDereverb +from vr import AudioPre, AudioPreDeEcho + +weight_uvr5_root = "tools/uvr5/uvr5_weights" +uvr5_names = [] +for name in os.listdir(weight_uvr5_root): + if name.endswith(".pth") or "onnx" in name: + uvr5_names.append(name.replace(".pth", "")) + +device=sys.argv[1] +is_half=sys.argv[2] +webui_port_uvr5=int(sys.argv[3]) +is_share=eval(sys.argv[4]) + +def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg, format0): + infos = [] + try: + inp_root = inp_root.strip(" ").strip('"').strip("\n").strip('"').strip(" ") + save_root_vocal = ( + save_root_vocal.strip(" ").strip('"').strip("\n").strip('"').strip(" ") + ) + save_root_ins = ( + save_root_ins.strip(" ").strip('"').strip("\n").strip('"').strip(" ") + ) + if model_name == "onnx_dereverb_By_FoxJoy": + from MDXNet import MDXNetDereverb + + pre_fun = MDXNetDereverb(15) + else: + func = AudioPre if "DeEcho" not in model_name else AudioPreDeEcho + pre_fun = func( + agg=int(agg), + model_path=os.path.join(weight_uvr5_root, model_name + ".pth"), + device=device, + is_half=is_half, + ) + if inp_root != "": + paths = [os.path.join(inp_root, name) for name in os.listdir(inp_root)] + else: + paths = [path.name for path in paths] + for path in paths: + inp_path = os.path.join(inp_root, path) + if(os.path.isfile(inp_path)==False):continue + need_reformat = 1 + done = 0 + try: + info = ffmpeg.probe(inp_path, cmd="ffprobe") + if ( + info["streams"][0]["channels"] == 2 + and info["streams"][0]["sample_rate"] == "44100" + ): + need_reformat = 0 + pre_fun._path_audio_( + inp_path, save_root_ins, save_root_vocal, format0 + ) + done = 1 + except: + need_reformat = 1 + traceback.print_exc() + if need_reformat == 1: + tmp_path = "%s/%s.reformatted.wav" % ( + os.path.join(os.environ["TEMP"]), + os.path.basename(inp_path), + ) + os.system( + "ffmpeg -i %s -vn -acodec pcm_s16le -ac 2 -ar 44100 %s -y" + % (inp_path, tmp_path) + ) + inp_path = tmp_path + try: + if done == 0: + pre_fun._path_audio_( + inp_path, save_root_ins, save_root_vocal, format0 + ) + infos.append("%s->Success" % (os.path.basename(inp_path))) + yield "\n".join(infos) + except: + infos.append( + "%s->%s" % (os.path.basename(inp_path), traceback.format_exc()) + ) + yield "\n".join(infos) + except: + infos.append(traceback.format_exc()) + yield "\n".join(infos) + finally: + try: + if model_name == "onnx_dereverb_By_FoxJoy": + del pre_fun.pred.model + del pre_fun.pred.model_ + else: + del pre_fun.model + del pre_fun + except: + traceback.print_exc() + print("clean_empty_cache") + if torch.cuda.is_available(): + torch.cuda.empty_cache() + yield "\n".join(infos) + +with gr.Blocks(title="UVR5 WebUI") as app: + gr.Markdown( + value= + i18n("本软件以MIT协议开源, 作者不对软件具备任何控制力, 使用软件者、传播软件导出的声音者自负全责.
如不认可该条款, 则不能使用或引用软件包内任何代码和文件. 详见根目录LICENSE.") + ) + with gr.Tabs(): + with gr.TabItem(i18n("伴奏人声分离&去混响&去回声")): + with gr.Group(): + gr.Markdown( + value=i18n( + "人声伴奏分离批量处理， 使用UVR5模型。
合格的文件夹路径格式举例： E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华测试样例(去文件管理器地址栏拷就行了)。
模型分为三类：
1、保留人声：不带和声的音频选这个，对主人声保留比HP5更好。内置HP2和HP3两个模型，HP3可能轻微漏伴奏但对主人声保留比HP2稍微好一丁点；
2、仅保留主人声：带和声的音频选这个，对主人声可能有削弱。内置HP5一个模型；
3、去混响、去延迟模型（by FoxJoy）：
  (1)MDX-Net(onnx_dereverb):对于双通道混响是最好的选择，不能去除单通道混响；
 (234)DeEcho:去除延迟效果。Aggressive比Normal去除得更彻底，DeReverb额外去除混响，可去除单声道混响，但是对高频重的板式混响去不干净。
去混响/去延迟，附：
1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；
2、MDX-Net-Dereverb模型挺慢的；
3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。" + ) + ) + with gr.Row(): + with gr.Column(): + dir_wav_input = gr.Textbox( + label=i18n("输入待处理音频文件夹路径"), + placeholder="C:\\Users\\Desktop\\todo-songs", + ) + wav_inputs = gr.File( + file_count="multiple", label=i18n("也可批量输入音频文件, 二选一, 优先读文件夹") + ) + with gr.Column(): + model_choose = gr.Dropdown(label=i18n("模型"), choices=uvr5_names) + agg = gr.Slider( + minimum=0, + maximum=20, + step=1, + label=i18n("人声提取激进程度"), + value=10, + interactive=True, + visible=False, # 先不开放调整 + ) + opt_vocal_root = gr.Textbox( + label=i18n("指定输出主人声文件夹"), value="output/uvr5_opt" + ) + opt_ins_root = gr.Textbox( + label=i18n("指定输出非主人声文件夹"), value="output/uvr5_opt" + ) + format0 = gr.Radio( + label=i18n("导出文件格式"), + choices=["wav", "flac", "mp3", "m4a"], + value="flac", + interactive=True, + ) + but2 = gr.Button(i18n("转换"), variant="primary") + vc_output4 = gr.Textbox(label=i18n("输出信息")) + but2.click( + uvr, + [ + model_choose, + dir_wav_input, + opt_vocal_root, + wav_inputs, + opt_ins_root, + agg, + format0, + ], + [vc_output4], + api_name="uvr_convert", + ) +app.queue(concurrency_count=511, max_size=1022).launch( + server_name="0.0.0.0", + inbrowser=True, + share=is_share, + server_port=webui_port_uvr5, + quiet=True, +)