From 158f85c74521c53e4af14c7f5d5e216c1367d5cc Mon Sep 17 00:00:00 2001 From: KakaruHayate Date: Sat, 30 Mar 2024 21:41:51 +0800 Subject: [PATCH] Update inference_webui.py --- GPT_SoVITS/inference_webui.py | 434 ++++++++++++++++++++-------------- 1 file changed, 263 insertions(+), 171 deletions(-) diff --git a/GPT_SoVITS/inference_webui.py b/GPT_SoVITS/inference_webui.py index caedc564..237d1b50 100644 --- a/GPT_SoVITS/inference_webui.py +++ b/GPT_SoVITS/inference_webui.py @@ -1,25 +1,47 @@ -import os,re,logging +''' +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 +import torch + +device = "cpu" + +try: + import torch_musa + use_torch_musa = True +except ImportError: + use_torch_musa = False +if use_torch_musa: + if "_MUSA_VISIBLE_DEVICES" in os.environ: + os.environ["MUSA_VISIBLE_DEVICES"] = os.environ["_MUSA_VISIBLE_DEVICES"] + if torch.musa.is_available(): + device = "musa" if os.path.exists("./gweight.txt"): - with open("./gweight.txt", 'r',encoding="utf-8") as file: + with open("./gweight.txt", 'r', encoding="utf-8") as file: gweight_data = file.read() gpt_path = os.environ.get( - "gpt_path", gweight_data) + "gpt_path", gweight_data) else: gpt_path = os.environ.get( - "gpt_path", "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt") + "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: + with open("./sweight.txt", 'r', encoding="utf-8") as file: sweight_data = file.read() sovits_path = os.environ.get("sovits_path", sweight_data) else: @@ -29,24 +51,25 @@ else: # ) # sovits_path = os.environ.get("sovits_path", "pretrained_models/s2G488k.pth") cnhubert_base_path = os.environ.get( - "cnhubert_base_path", "pretrained_models/chinese-hubert-base" + "cnhubert_base_path", "GPT_SoVITS/pretrained_models/chinese-hubert-base" ) bert_path = os.environ.get( - "bert_path", "pretrained_models/chinese-roberta-wwm-ext-large" + "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) +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")) +is_half = eval(os.environ.get("is_half", "True")) and torch.cuda.is_available() import gradio as gr from transformers import AutoModelForMaskedLM, AutoTokenizer import numpy as np -import librosa,torch +import librosa from feature_extractor import cnhubert -cnhubert.cnhubert_base_path=cnhubert_base_path + +cnhubert.cnhubert_base_path = cnhubert_base_path from module.models import SynthesizerTrn from AR.models.t2s_lightning_module import Text2SemanticLightningModule @@ -56,16 +79,13 @@ 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。 +# 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) @@ -74,6 +94,7 @@ if is_half == True: else: bert_model = bert_model.to(device) + def get_bert_feature(text, word2ph): with torch.no_grad(): inputs = tokenizer(text, return_tensors="pt") @@ -89,6 +110,7 @@ def get_bert_feature(text, word2ph): 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) @@ -123,10 +145,11 @@ if is_half == True: 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"] + 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( @@ -135,7 +158,7 @@ def change_sovits_weights(sovits_path): n_speakers=hps.data.n_speakers, **hps.model ) - if("pretrained"not in sovits_path): + if ("pretrained" not in sovits_path): del vq_model.enc_q if is_half == True: vq_model = vq_model.half().to(device) @@ -143,11 +166,15 @@ def change_sovits_weights(sovits_path): 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) + 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 + global hz, max_sec, t2s_model, config hz = 50 dict_s1 = torch.load(gpt_path, map_location="cpu") config = dict_s1["config"] @@ -160,9 +187,12 @@ def change_gpt_weights(gpt_path): 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) + 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) @@ -179,43 +209,26 @@ def get_spepc(hps, filename): return spec -dict_language={ - i18n("中文"):"zh", - i18n("英文"):"en", - i18n("日文"):"ja" +dict_language = { + 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 } -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 - +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)), @@ -225,54 +238,112 @@ def get_bert_inf(phones, word2ph, norm_text, language): 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 +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_phones_and_bert(text,language): + if language in {"en","all_zh","all_ja"}: + language = language.replace("all_","") + if language == "en": + LangSegment.setfilters(["en"]) + formattext = " ".join(tmp["text"] for tmp in LangSegment.getTexts(text)) 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) + # 因无法区别中日文汉字,以用户输入为准 + formattext = text + while " " in formattext: + formattext = formattext.replace(" ", " ") + phones, word2ph, norm_text = clean_text_inf(formattext, 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) + elif language in {"zh", "ja","auto"}: + textlist=[] + langlist=[] + LangSegment.setfilters(["zh","ja","en","ko"]) + if language == "auto": + for tmp in LangSegment.getTexts(text): + if tmp["lang"] == "ko": + langlist.append("zh") + textlist.append(tmp["text"]) + else: + langlist.append(tmp["lang"]) + textlist.append(tmp["text"]) + else: + for tmp in LangSegment.getTexts(text): + if tmp["lang"] == "en": + langlist.append(tmp["lang"]) + else: + # 因无法区别中日文汉字,以用户输入为准 + langlist.append(language) + textlist.append(tmp["text"]) + print(textlist) + print(langlist) + phones_list = [] + bert_list = [] + norm_text_list = [] + for i in range(len(textlist)): + lang = langlist[i] + phones, word2ph, norm_text = clean_text_inf(textlist[i], lang) + bert = get_bert_inf(phones, word2ph, norm_text, lang) + phones_list.append(phones) + norm_text_list.append(norm_text) + bert_list.append(bert) + bert = torch.cat(bert_list, dim=1) + phones = sum(phones_list, []) + norm_text = ''.join(norm_text_list) - return phones, word2ph, norm_text + return phones,bert.to(dtype),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) +def merge_short_text_in_array(texts, threshold): + if (len(texts)) < 2: + return texts + result = [] + text = "" + for ele in texts: + text += ele + if len(text) >= threshold: + result.append(text) + text = "" + if (len(text) > 0): + if len(result) == 0: + result.append(text) + else: + result[len(result) - 1] += text + return result - return bert - -#i18n("不切"),i18n("凑五句一切"),i18n("凑50字一切"),i18n("按中文句号。切"),i18n("按英文句号.切") -def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language,how_to_cut=i18n("不切")): +def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切"), top_k=20, top_p=0.6, temperature=0.6, ref_free = False): + if prompt_text is None or len(prompt_text) == 0: + ref_free = True t0 = ttime() - prompt_text = prompt_text.strip("\n") - prompt_language, text = prompt_language, text.strip("\n") + prompt_language = dict_language[prompt_language] + text_language = dict_language[text_language] + if not ref_free: + prompt_text = prompt_text.strip("\n") + if (prompt_text[-1] not in splits): prompt_text += "。" if prompt_language != "en" else "." + print(i18n("实际输入的参考文本:"), prompt_text) + text = text.strip("\n") + if (text[0] not in splits and len(get_first(text)) < 4): text = "。" + text if text_language != "en" else "." + text + + print(i18n("实际输入的目标文本:"), text) 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: @@ -281,52 +352,51 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, else: wav16k = wav16k.to(device) zero_wav_torch = zero_wav_torch.to(device) - wav16k=torch.cat([wav16k,zero_wav_torch]) + 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("凑五句一切")):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") + 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) + elif (how_to_cut == i18n("按标点符号切")): + text = cut5(text) + while "\n\n" in text: + text = text.replace("\n\n", "\n") + print(i18n("实际输入的目标文本(切句后):"), text) + texts = text.split("\n") + texts = merge_short_text_in_array(texts, 5) 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) - + if not ref_free: + phones1,bert1,norm_text1=get_phones_and_bert(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) + 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) + phones2,bert2,norm_text2=get_phones_and_bert(text, text_language) + print(i18n("前端处理后的文本(每句):"), norm_text2) + if not ref_free: + bert = torch.cat([bert1, bert2], 1) + all_phoneme_ids = torch.LongTensor(phones1+phones2).to(device).unsqueeze(0) else: - phones2, word2ph2, norm_text2 = nonen_clean_text_inf(text, text_language) + bert = bert2 + all_phoneme_ids = torch.LongTensor(phones2).to(device).unsqueeze(0) - 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) @@ -336,10 +406,12 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, pred_semantic, idx = t2s_model.model.infer_panel( all_phoneme_ids, all_phoneme_len, - prompt, + None if ref_free else prompt, bert, # prompt_phone_len=ph_offset, - top_k=config["inference"]["top_k"], + top_k=top_k, + top_p=top_p, + temperature=temperature, early_stop_num=hz * max_sec, ) t3 = ttime() @@ -357,10 +429,12 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, vq_model.decode( pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refer ) - .detach() - .cpu() - .numpy()[0, 0] + .detach() + .cpu() + .numpy()[0, 0] ) ###试试重建不带上prompt部分 + max_audio=np.abs(audio).max()#简单防止16bit爆音 + if max_audio>1:audio/=max_audio audio_opt.append(audio) audio_opt.append(zero_wav) t4 = ttime() @@ -370,23 +444,6 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, ) -splits = { - "，", - "。", - "？", - "！", - ",", - ".", - "?", - "!", - "~", - ":", - "：", - "—", - "…", -} # 不考虑省略号 - - def split(todo_text): todo_text = todo_text.replace("……", "。").replace("——", "，") if todo_text[-1] not in splits: @@ -409,12 +466,12 @@ def split(todo_text): def cut1(inp): inp = inp.strip("\n") inps = split(inp) - split_idx = list(range(0, len(inps), 5)) + 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]])) + opts.append("".join(inps[split_idx[idx]: split_idx[idx + 1]])) else: opts = [inp] return "\n".join(opts) @@ -424,7 +481,7 @@ def cut2(inp): inp = inp.strip("\n") inps = split(inp) if len(inps) < 2: - return [inp] + return inp opts = [] summ = 0 tmp_str = "" @@ -437,7 +494,8 @@ def cut2(inp): tmp_str = "" if tmp_str != "": opts.append(tmp_str) - if len(opts[-1]) < 50: ##如果最后一个太短了，和前一个合一起 + # 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) @@ -445,10 +503,28 @@ def cut2(inp): def cut3(inp): inp = inp.strip("\n") - return "\n".join(["%s。" % item for item in inp.strip("。").split("。")]) + 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(".")]) + return "\n".join(["%s" % item for item in inp.strip(".").split(".")]) + + +# contributed by https://github.com/AI-Hobbyist/GPT-SoVITS/blob/main/GPT_SoVITS/inference_webui.py +def cut5(inp): + # if not re.search(r'[^\w\s]', inp[-1]): + # inp += '。' + inp = inp.strip("\n") + punds = r'[,.;?!、，。？！;：…]' + items = re.split(f'({punds})', inp) + mergeitems = ["".join(group) for group in zip(items[::2], items[1::2])] + # 在句子不存在符号或句尾无符号的时候保证文本完整 + if len(items)%2 == 1: + mergeitems.append(items[-1]) + opt = "\n".join(mergeitems) + return opt + 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 @@ -457,25 +533,31 @@ def custom_sort_key(s): 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"} + 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) + -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)) + 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() + 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() with gr.Blocks(title="GPT-SoVITS WebUI") as app: gr.Markdown( @@ -484,53 +566,63 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app: with gr.Group(): gr.Markdown(value=i18n("模型切换")) with gr.Row(): - GPT_dropdown = gr.Dropdown(label=i18n("GPT模型列表"), choices=sorted(GPT_names, key=custom_sort_key), value=gpt_path,interactive=True) - SoVITS_dropdown = gr.Dropdown(label=i18n("SoVITS模型列表"), choices=sorted(SoVITS_names, key=custom_sort_key), value=sovits_path,interactive=True) + GPT_dropdown = gr.Dropdown(label=i18n("GPT模型列表"), choices=sorted(GPT_names, key=custom_sort_key), value=gpt_path, interactive=True) + SoVITS_dropdown = gr.Dropdown(label=i18n("SoVITS模型列表"), choices=sorted(SoVITS_names, key=custom_sort_key), value=sovits_path, interactive=True) refresh_button = gr.Button(i18n("刷新模型路径"), variant="primary") refresh_button.click(fn=change_choices, inputs=[], outputs=[SoVITS_dropdown, GPT_dropdown]) - SoVITS_dropdown.change(change_sovits_weights,[SoVITS_dropdown],[]) - GPT_dropdown.change(change_gpt_weights,[GPT_dropdown],[]) + SoVITS_dropdown.change(change_sovits_weights, [SoVITS_dropdown], []) + GPT_dropdown.change(change_gpt_weights, [GPT_dropdown], []) gr.Markdown(value=i18n("*请上传并填写参考信息")) with gr.Row(): - inp_ref = gr.Audio(label=i18n("请上传参考音频"), type="filepath") - prompt_text = gr.Textbox(label=i18n("参考音频的文本"), value="") + inp_ref = gr.Audio(label=i18n("请上传3~10秒内参考音频，超过会报错！"), type="filepath") + with gr.Column(): + ref_text_free = gr.Checkbox(label=i18n("开启无参考文本模式。不填参考文本亦相当于开启。"), value=False, interactive=True, show_label=True) + gr.Markdown(i18n("使用无参考文本模式时建议使用微调的GPT，听不清参考音频说的啥(不晓得写啥)可以开，开启后无视填写的参考文本。")) + prompt_text = gr.Textbox(label=i18n("参考音频的文本"), value="") prompt_language = gr.Dropdown( - label=i18n("参考音频的语种"),choices=[i18n("中文"),i18n("英文"),i18n("日文")],value=i18n("中文") + label=i18n("参考音频的语种"), 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") ) - gr.Markdown(value=i18n("*请填写需要合成的目标文本。中英混合选中文，日英混合选日文，中日混合暂不支持，非目标语言文本自动遗弃。")) + gr.Markdown(value=i18n("*请填写需要合成的目标文本和语种模式")) 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"), - choices=[i18n("不切"),i18n("凑五句一切"),i18n("凑50字一切"),i18n("按中文句号。切"),i18n("按英文句号.切"),], - value=i18n("凑50字一切"), + choices=[i18n("不切"), i18n("凑四句一切"), i18n("凑50字一切"), i18n("按中文句号。切"), i18n("按英文句号.切"), i18n("按标点符号切"), ], + value=i18n("凑四句一切"), interactive=True, ) + with gr.Row(): + gr.Markdown(value=i18n("gpt采样参数(无参考文本时不要太低)：")) + top_k = gr.Slider(minimum=1,maximum=100,step=1,label=i18n("top_k"),value=5,interactive=True) + top_p = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("top_p"),value=1,interactive=True) + temperature = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("temperature"),value=1,interactive=True) inference_button = gr.Button(i18n("合成语音"), variant="primary") output = gr.Audio(label=i18n("输出的语音")) inference_button.click( get_tts_wav, - [inp_ref, prompt_text, prompt_language, text, text_language,how_to_cut], + [inp_ref, prompt_text, prompt_language, text, text_language, how_to_cut, top_k, top_p, temperature, ref_text_free], [output], ) gr.Markdown(value=i18n("文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。")) with gr.Row(): - text_inp = gr.Textbox(label=i18n("需要合成的切分前文本"),value="") - button1 = gr.Button(i18n("凑五句一切"), variant="primary") + text_inp = gr.Textbox(label=i18n("需要合成的切分前文本"), value="") + button1 = gr.Button(i18n("凑四句一切"), variant="primary") button2 = gr.Button(i18n("凑50字一切"), variant="primary") button3 = gr.Button(i18n("按中文句号。切"), variant="primary") button4 = gr.Button(i18n("按英文句号.切"), variant="primary") + button5 = gr.Button(i18n("按标点符号切"), variant="primary") text_opt = gr.Textbox(label=i18n("切分后文本"), value="") button1.click(cut1, [text_inp], [text_opt]) button2.click(cut2, [text_inp], [text_opt]) button3.click(cut3, [text_inp], [text_opt]) button4.click(cut4, [text_inp], [text_opt]) - gr.Markdown(value=i18n("后续将支持混合语种编码文本输入。")) + button5.click(cut5, [text_inp], [text_opt]) + gr.Markdown(value=i18n("后续将支持转音素、手工修改音素、语音合成分步执行。")) app.queue(concurrency_count=511, max_size=1022).launch( server_name="0.0.0.0",