''' 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, sys now_dir = os.getcwd() sys.path.append(now_dir) 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 import torch 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")) and torch.cuda.is_available() gpt_path = os.environ.get("gpt_path", None) sovits_path = os.environ.get("sovits_path", None) cnhubert_base_path = os.environ.get("cnhubert_base_path", None) bert_path = os.environ.get("bert_path", None) import gradio as gr from TTS_infer_pack.TTS import TTS, TTS_Config from TTS_infer_pack.text_segmentation_method import get_method 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" 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 } cut_method = { i18n("不切"):"cut0", i18n("凑四句一切"): "cut1", i18n("凑50字一切"): "cut2", i18n("按中文句号。切"): "cut3", i18n("按英文句号.切"): "cut4", i18n("按标点符号切"): "cut5", } tts_config = TTS_Config("GPT_SoVITS/configs/tts_infer.yaml") tts_config.device = device tts_config.is_half = is_half if gpt_path is not None: tts_config.t2s_weights_path = gpt_path if sovits_path is not None: tts_config.vits_weights_path = sovits_path if cnhubert_base_path is not None: tts_config.cnhuhbert_base_path = cnhubert_base_path if bert_path is not None: tts_config.bert_base_path = bert_path print(tts_config) tts_pipline = TTS(tts_config) gpt_path = tts_config.t2s_weights_path sovits_path = tts_config.vits_weights_path def inference(text, text_lang, ref_audio_path, prompt_text, prompt_lang, top_k, top_p, temperature, text_split_method, batch_size, speed_factor, ref_text_free, split_bucket,fragment_interval, seed, ): inputs={ "text": text, "text_lang": dict_language[text_lang], "ref_audio_path": ref_audio_path, "prompt_text": prompt_text if not ref_text_free else "", "prompt_lang": dict_language[prompt_lang], "top_k": top_k, "top_p": top_p, "temperature": temperature, "text_split_method": cut_method[text_split_method], "batch_size":int(batch_size), "speed_factor":float(speed_factor), "split_bucket":split_bucket, "return_fragment":False, "fragment_interval":fragment_interval, "seed":seed, } for item in tts_pipline.run(inputs): yield item 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() with gr.Blocks(title="GPT-SoVITS 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.Column(): # 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) 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(tts_pipline.init_vits_weights, [SoVITS_dropdown], []) GPT_dropdown.change(tts_pipline.init_t2s_weights, [GPT_dropdown], []) with gr.Row(): with gr.Column(): gr.Markdown(value=i18n("*请上传并填写参考信息")) inp_ref = gr.Audio(label=i18n("请上传3~10秒内参考音频,超过会报错!"), type="filepath") prompt_text = gr.Textbox(label=i18n("参考音频的文本"), value="", lines=2) with gr.Row(): prompt_language = gr.Dropdown( 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") ) 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。")) with gr.Column(): gr.Markdown(value=i18n("*请填写需要合成的目标文本和语种模式")) text = gr.Textbox(label=i18n("需要合成的文本"), value="", lines=16, max_lines=16) text_language = gr.Dropdown( 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") ) with gr.Group(): gr.Markdown(value=i18n("推理设置")) with gr.Row(): with gr.Column(): batch_size = gr.Slider(minimum=1,maximum=200,step=1,label=i18n("batch_size"),value=20,interactive=True) fragment_interval = gr.Slider(minimum=0.01,maximum=1,step=0.01,label=i18n("分段间隔(秒)"),value=0.3,interactive=True) speed_factor = gr.Slider(minimum=0.25,maximum=4,step=0.05,label="speed_factor",value=1.0,interactive=True) 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) with gr.Column(): how_to_cut = gr.Radio( label=i18n("怎么切"), 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(): split_bucket = gr.Checkbox(label=i18n("数据分桶(可能会降低一点计算量,选就对了)"), value=True, interactive=True, show_label=True) seed = gr.Number(label=i18n("随机种子"),value=-1) # with gr.Column(): output = gr.Audio(label=i18n("输出的语音")) with gr.Row(): inference_button = gr.Button(i18n("合成语音"), variant="primary") stop_infer = gr.Button(i18n("终止合成"), variant="primary") inference_button.click( inference, [ text,text_language, inp_ref, prompt_text, prompt_language, top_k, top_p, temperature, how_to_cut, batch_size, speed_factor, ref_text_free, split_bucket,fragment_interval, seed ], [output],api_name="GetVoice" ) stop_infer.click(tts_pipline.stop, [], []) with gr.Group(): 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="", lines=4) with gr.Column(): _how_to_cut = gr.Radio( label=i18n("怎么切"), 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, ) cut_text= gr.Button(i18n("切分"), variant="primary") def to_cut(text_inp, how_to_cut): if len(text_inp.strip()) == 0 or text_inp==[]: return "" method = get_method(cut_method[how_to_cut]) return method(text_inp) text_opt = gr.Textbox(label=i18n("切分后文本"), value="", lines=4) cut_text.click(to_cut, [text_inp, _how_to_cut], [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().launch( server_name="0.0.0.0", inbrowser=True, share=is_share, server_port=infer_ttswebui, quiet=True, )