kwau/sovits-isla

1import glob2import json3import logging4import os5import re6import subprocess7import sys8import time9import traceback10from itertools import chain11from pathlib import Path12 13# os.system("wget -P cvec/ https://huggingface.co/spaces/innnky/nanami/resolve/main/checkpoint_best_legacy_500.pt")14import gradio as gr15import librosa16import numpy as np17import soundfile18import torch19 20from compress_model import removeOptimizer21from edgetts.tts_voices import SUPPORTED_LANGUAGES22from inference.infer_tool import Svc23from utils import mix_model24 25logging.getLogger('numba').setLevel(logging.WARNING)26logging.getLogger('markdown_it').setLevel(logging.WARNING)27logging.getLogger('urllib3').setLevel(logging.WARNING)28logging.getLogger('matplotlib').setLevel(logging.WARNING)29logging.getLogger('multipart').setLevel(logging.WARNING)30 31model = None32spk = None33debug = False34 35local_model_root = './trained'36 37cuda = {}38if torch.cuda.is_available():39    for i in range(torch.cuda.device_count()):40        device_name = torch.cuda.get_device_properties(i).name41        cuda[f"CUDA:{i} {device_name}"] = f"cuda:{i}"42 43def upload_mix_append_file(files,sfiles):44    try:45        if(sfiles is None):46            file_paths = [file.name for file in files]47        else:48            file_paths = [file.name for file in chain(files,sfiles)]49        p = {file:100 for file in file_paths}50        return file_paths,mix_model_output1.update(value=json.dumps(p,indent=2))51    except Exception as e:52        if debug:53            traceback.print_exc()54        raise gr.Error(e)55 56def mix_submit_click(js,mode):57    try:58        assert js.lstrip()!=""59        modes = {"凸组合":0, "线性组合":1}60        mode = modes[mode]61        data = json.loads(js)62        data = list(data.items())63        model_path,mix_rate = zip(*data)64        path = mix_model(model_path,mix_rate,mode)65        return f"成功，文件被保存在了{path}"66    except Exception as e:67        if debug:68            traceback.print_exc()69        raise gr.Error(e)70 71def updata_mix_info(files):72    try:73        if files is None :74            return mix_model_output1.update(value="")75        p = {file.name:100 for file in files}76        return mix_model_output1.update(value=json.dumps(p,indent=2))77    except Exception as e:78        if debug:79            traceback.print_exc()80        raise gr.Error(e)81 82def modelAnalysis(model_path,config_path,cluster_model_path,device,enhance,diff_model_path,diff_config_path,only_diffusion,use_spk_mix,local_model_enabled,local_model_selection):83    global model84    try:85        device = cuda[device] if "CUDA" in device else device86        cluster_filepath = os.path.split(cluster_model_path.name) if cluster_model_path is not None else "no_cluster"87        # get model and config path88        if (local_model_enabled):89            # local path90            model_path = glob.glob(os.path.join(local_model_selection, '*.pth'))[0]91            config_path = glob.glob(os.path.join(local_model_selection, '*.json'))[0]92        else:93            # upload from webpage94            model_path = model_path.name95            config_path = config_path.name96        fr = ".pkl" in cluster_filepath[1]97        model = Svc(model_path,98                config_path,99                device=device if device != "Auto" else None,100                cluster_model_path = cluster_model_path.name if cluster_model_path is not None else "",101                nsf_hifigan_enhance=enhance,102                diffusion_model_path = diff_model_path.name if diff_model_path is not None else "",103                diffusion_config_path = diff_config_path.name if diff_config_path is not None else "",104                shallow_diffusion = True if diff_model_path is not None else False,105                only_diffusion = only_diffusion,106                spk_mix_enable = use_spk_mix,107                feature_retrieval = fr108                )109        spks = list(model.spk2id.keys())110        device_name = torch.cuda.get_device_properties(model.dev).name if "cuda" in str(model.dev) else str(model.dev)111        msg = f"成功加载模型到设备{device_name}上\n"112        if cluster_model_path is None:113            msg += "未加载聚类模型或特征检索模型\n"114        elif fr:115            msg += f"特征检索模型{cluster_filepath[1]}加载成功\n"116        else:117            msg += f"聚类模型{cluster_filepath[1]}加载成功\n"118        if diff_model_path is None:119            msg += "未加载扩散模型\n"120        else:121            msg += f"扩散模型{diff_model_path.name}加载成功\n"122        msg += "当前模型的可用音色：\n"123        for i in spks:124            msg += i + " "125        return sid.update(choices = spks,value=spks[0]), msg126    except Exception as e:127        if debug:128            traceback.print_exc()129        raise gr.Error(e)130 131    132def modelUnload():133    global model134    if model is None:135        return sid.update(choices = [],value=""),"没有模型需要卸载!"136    else:137        model.unload_model()138        model = None139        torch.cuda.empty_cache()140        return sid.update(choices = [],value=""),"模型卸载完毕!"141    142def vc_infer(output_format, sid, audio_path, truncated_basename, vc_transform, auto_f0, cluster_ratio, slice_db, noise_scale, pad_seconds, cl_num, lg_num, lgr_num, f0_predictor, enhancer_adaptive_key, cr_threshold, k_step, use_spk_mix, second_encoding, loudness_envelope_adjustment):143    global model144    _audio = model.slice_inference(145        audio_path,146        sid,147        vc_transform,148        slice_db,149        cluster_ratio,150        auto_f0,151        noise_scale,152        pad_seconds,153        cl_num,154        lg_num,155        lgr_num,156        f0_predictor,157        enhancer_adaptive_key,158        cr_threshold,159        k_step,160        use_spk_mix,161        second_encoding,162        loudness_envelope_adjustment163    )  164    model.clear_empty()165    #构建保存文件的路径，并保存到results文件夹内166    str(int(time.time()))167    if not os.path.exists("results"):168        os.makedirs("results")169    key = "auto" if auto_f0 else f"{int(vc_transform)}key"170    cluster = "_" if cluster_ratio == 0 else f"_{cluster_ratio}_"171    isdiffusion = "sovits"172    if model.shallow_diffusion:173        isdiffusion = "sovdiff"174 175    if model.only_diffusion:176        isdiffusion = "diff"177    178    output_file_name = 'result_'+truncated_basename+f'_{sid}_{key}{cluster}{isdiffusion}.{output_format}'179    output_file = os.path.join("results", output_file_name)180    soundfile.write(output_file, _audio, model.target_sample, format=output_format)181    return output_file182 183def vc_fn(sid, input_audio, output_format, vc_transform, auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,f0_predictor,enhancer_adaptive_key,cr_threshold,k_step,use_spk_mix,second_encoding,loudness_envelope_adjustment):184    global model185    try:186        if input_audio is None:187            return "You need to upload an audio", None188        if model is None:189            return "You need to upload an model", None190        if getattr(model, 'cluster_model', None) is None and model.feature_retrieval is False:191            if cluster_ratio != 0:192                return "You need to upload an cluster model or feature retrieval model before assigning cluster ratio!", None193        #print(input_audio)    194        audio, sampling_rate = soundfile.read(input_audio)195        #print(audio.shape,sampling_rate)196        if np.issubdtype(audio.dtype, np.integer):197            audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)198        #print(audio.dtype)199        if len(audio.shape) > 1:200            audio = librosa.to_mono(audio.transpose(1, 0))201        # 未知原因Gradio上传的filepath会有一个奇怪的固定后缀，这里去掉202        truncated_basename = Path(input_audio).stem[:-6]203        processed_audio = os.path.join("raw", f"{truncated_basename}.wav")204        soundfile.write(processed_audio, audio, sampling_rate, format="wav")205        output_file = vc_infer(output_format, sid, processed_audio, truncated_basename, vc_transform, auto_f0, cluster_ratio, slice_db, noise_scale, pad_seconds, cl_num, lg_num, lgr_num, f0_predictor, enhancer_adaptive_key, cr_threshold, k_step, use_spk_mix, second_encoding, loudness_envelope_adjustment)206 207        return "Success", output_file208    except Exception as e:209        if debug:210            traceback.print_exc()211        raise gr.Error(e)212 213def text_clear(text):214    return re.sub(r"[\n\,\(\) ]", "", text)215 216def vc_fn2(_text, _lang, _gender, _rate, _volume, sid, output_format, vc_transform, auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,f0_predictor,enhancer_adaptive_key,cr_threshold, k_step,use_spk_mix,second_encoding,loudness_envelope_adjustment):217    global model218    try:219        if model is None:220            return "You need to upload an model", None221        if getattr(model, 'cluster_model', None) is None and model.feature_retrieval is False:222            if cluster_ratio != 0:223                return "You need to upload an cluster model or feature retrieval model before assigning cluster ratio!", None224        _rate = f"+{int(_rate*100)}%" if _rate >= 0 else f"{int(_rate*100)}%"225        _volume = f"+{int(_volume*100)}%" if _volume >= 0 else f"{int(_volume*100)}%"226        if _lang == "Auto":227            _gender = "Male" if _gender == "男" else "Female"228            subprocess.run([sys.executable, "edgetts/tts.py", _text, _lang, _rate, _volume, _gender])229        else:230            subprocess.run([sys.executable, "edgetts/tts.py", _text, _lang, _rate, _volume])231        target_sr = 44100232        y, sr = librosa.load("tts.wav")233        resampled_y = librosa.resample(y, orig_sr=sr, target_sr=target_sr)234        soundfile.write("tts.wav", resampled_y, target_sr, subtype = "PCM_16")235        input_audio = "tts.wav"236        #audio, _ = soundfile.read(input_audio)237        output_file_path = vc_infer(output_format, sid, input_audio, "tts", vc_transform, auto_f0, cluster_ratio, slice_db, noise_scale, pad_seconds, cl_num, lg_num, lgr_num, f0_predictor, enhancer_adaptive_key, cr_threshold, k_step, use_spk_mix, second_encoding, loudness_envelope_adjustment)238        os.remove("tts.wav")239        return "Success", output_file_path240    except Exception as e:241        if debug: traceback.print_exc()  # noqa: E701242        raise gr.Error(e)243 244def model_compression(_model):245    if _model == "":246        return "请先选择要压缩的模型"247    else:248        model_path = os.path.split(_model.name)249        filename, extension = os.path.splitext(model_path[1])250        output_model_name = f"{filename}_compressed{extension}"251        output_path = os.path.join(os.getcwd(), output_model_name)252        removeOptimizer(_model.name, output_path)253        return f"模型已成功被保存在了{output_path}"254 255def scan_local_models():256    res = []257    candidates = glob.glob(os.path.join(local_model_root, '**', '*.json'), recursive=True)258    candidates = set([os.path.dirname(c) for c in candidates])259    for candidate in candidates:260        jsons = glob.glob(os.path.join(candidate, '*.json'))261        pths = glob.glob(os.path.join(candidate, '*.pth'))262        if (len(jsons) == 1 and len(pths) == 1):263            # must contain exactly one json and one pth file264            res.append(candidate)265    return res266 267def local_model_refresh_fn():268    choices = scan_local_models()269    return gr.Dropdown.update(choices=choices)270 271def debug_change():272    global debug273    debug = debug_button.value274 275with gr.Blocks(276    theme=gr.themes.Base(277        primary_hue = gr.themes.colors.green,278        font=["Source Sans Pro", "Arial", "sans-serif"],279        font_mono=['JetBrains mono', "Consolas", 'Courier New']280    ),281) as app:282    with gr.Tabs():283        with gr.TabItem("推理"):284            gr.Markdown(value="""285                So-vits-svc 4.0 推理 webui286                """)287            with gr.Row(variant="panel"):288                with gr.Column():289                    gr.Markdown(value="""290                        <font size=2> 模型设置</font>291                        """)292                    with gr.Tabs():293                        # invisible checkbox that tracks tab status294                        local_model_enabled = gr.Checkbox(value=False, visible=False)295                        with gr.TabItem('上传') as local_model_tab_upload:296                            with gr.Row():297                                model_path = gr.File(label="选择模型文件")298                                config_path = gr.File(label="选择配置文件")299                        with gr.TabItem('本地') as local_model_tab_local:300                            gr.Markdown(f'模型应当放置于{local_model_root}文件夹下')301                            local_model_refresh_btn = gr.Button('刷新本地模型列表')302                            local_model_selection = gr.Dropdown(label='选择模型文件夹', choices=[], interactive=True)303                    with gr.Row():304                        diff_model_path = gr.File(label="选择扩散模型文件")305                        diff_config_path = gr.File(label="选择扩散模型配置文件")306                    cluster_model_path = gr.File(label="选择聚类模型或特征检索文件（没有可以不选）")307                    device = gr.Dropdown(label="推理设备，默认为自动选择CPU和GPU", choices=["Auto",*cuda.keys(),"cpu"], value="Auto")308                    enhance = gr.Checkbox(label="是否使用NSF_HIFIGAN增强,该选项对部分训练集少的模型有一定的音质增强效果，但是对训练好的模型有反面效果，默认关闭", value=False)309                    only_diffusion = gr.Checkbox(label="是否使用全扩散推理，开启后将不使用So-VITS模型，仅使用扩散模型进行完整扩散推理，默认关闭", value=False)310                with gr.Column():311                    gr.Markdown(value="""312                        <font size=3>左侧文件全部选择完毕后(全部文件模块显示download)，点击“加载模型”进行解析：</font>313                        """)314                    model_load_button = gr.Button(value="加载模型", variant="primary")315                    model_unload_button = gr.Button(value="卸载模型", variant="primary")316                    sid = gr.Dropdown(label="音色（说话人）")317                    sid_output = gr.Textbox(label="Output Message")318 319 320            with gr.Row(variant="panel"):321                with gr.Column():322                    gr.Markdown(value="""323                        <font size=2> 推理设置</font>324                        """)325                    auto_f0 = gr.Checkbox(label="自动f0预测，配合聚类模型f0预测效果更好,会导致变调功能失效（仅限转换语音，歌声勾选此项会究极跑调）", value=False)326                    f0_predictor = gr.Dropdown(label="选择F0预测器,可选择crepe,pm,dio,harvest,rmvpe,默认为pm(注意：crepe为原F0使用均值滤波器)", choices=["pm","dio","harvest","crepe","rmvpe"], value="pm")327                    vc_transform = gr.Number(label="变调（整数，可以正负，半音数量，升高八度就是12）", value=0)328                    cluster_ratio = gr.Number(label="聚类模型/特征检索混合比例，0-1之间，0即不启用聚类/特征检索。使用聚类/特征检索能提升音色相似度，但会导致咬字下降（如果使用建议0.5左右）", value=0)329                    slice_db = gr.Number(label="切片阈值", value=-40)330                    output_format = gr.Radio(label="音频输出格式", choices=["wav", "flac", "mp3"], value = "wav")331                    noise_scale = gr.Number(label="noise_scale 建议不要动，会影响音质，玄学参数", value=0.4)332                    k_step = gr.Slider(label="浅扩散步数，只有使用了扩散模型才有效，步数越大越接近扩散模型的结果", value=100, minimum = 1, maximum = 1000)333                with gr.Column():334                    pad_seconds = gr.Number(label="推理音频pad秒数，由于未知原因开头结尾会有异响，pad一小段静音段后就不会出现", value=0.5)335                    cl_num = gr.Number(label="音频自动切片，0为不切片，单位为秒(s)", value=0)336                    lg_num = gr.Number(label="两端音频切片的交叉淡入长度，如果自动切片后出现人声不连贯可调整该数值，如果连贯建议采用默认值0，注意，该设置会影响推理速度，单位为秒/s", value=0)337                    lgr_num = gr.Number(label="自动音频切片后，需要舍弃每段切片的头尾。该参数设置交叉长度保留的比例，范围0-1,左开右闭", value=0.75)338                    enhancer_adaptive_key = gr.Number(label="使增强器适应更高的音域(单位为半音数)|默认为0", value=0)339                    cr_threshold = gr.Number(label="F0过滤阈值，只有启动crepe时有效. 数值范围从0-1. 降低该值可减少跑调概率，但会增加哑音", value=0.05)340                    loudness_envelope_adjustment = gr.Number(label="输入源响度包络替换输出响度包络融合比例，越靠近1越使用输出响度包络", value = 0)341                    second_encoding = gr.Checkbox(label = "二次编码，浅扩散前会对原始音频进行二次编码，玄学选项，效果时好时差，默认关闭", value=False)342                    use_spk_mix = gr.Checkbox(label = "动态声线融合", value = False, interactive = False)343            with gr.Tabs():344                with gr.TabItem("音频转音频"):345                    vc_input3 = gr.Audio(label="选择音频", type="filepath")346                    vc_submit = gr.Button("音频转换", variant="primary")347                with gr.TabItem("文字转音频"):348                    text2tts=gr.Textbox(label="在此输入要转译的文字。注意，使用该功能建议打开F0预测，不然会很怪")349                    with gr.Row():350                        tts_gender = gr.Radio(label = "说话人性别", choices = ["男","女"], value = "男")351                        tts_lang = gr.Dropdown(label = "选择语言，Auto为根据输入文字自动识别", choices=SUPPORTED_LANGUAGES, value = "Auto")352                        tts_rate = gr.Slider(label = "TTS语音变速（倍速相对值）", minimum = -1, maximum = 3, value = 0, step = 0.1)353                        tts_volume = gr.Slider(label = "TTS语音音量（相对值）", minimum = -1, maximum = 1.5, value = 0, step = 0.1)354                    vc_submit2 = gr.Button("文字转换", variant="primary")355            with gr.Row():356                with gr.Column():357                    vc_output1 = gr.Textbox(label="Output Message")358                with gr.Column():359                    vc_output2 = gr.Audio(label="Output Audio", interactive=False)360 361        with gr.TabItem("小工具/实验室特性"):362            gr.Markdown(value="""363                        <font size=2> So-vits-svc 4.0 小工具/实验室特性</font>364                        """)365            with gr.Tabs():366                with gr.TabItem("静态声线融合"):367                    gr.Markdown(value="""368                        <font size=2> 介绍:该功能可以将多个声音模型合成为一个声音模型(多个模型参数的凸组合或线性组合)，从而制造出现实中不存在的声线 369                                          注意：370                                          1.该功能仅支持单说话人的模型371                                          2.如果强行使用多说话人模型，需要保证多个模型的说话人数量相同，这样可以混合同一个SpaekerID下的声音372                                          3.保证所有待混合模型的config.json中的model字段是相同的373                                          4.输出的混合模型可以使用待合成模型的任意一个config.json，但聚类模型将不能使用374                                          5.批量上传模型的时候最好把模型放到一个文件夹选中后一起上传375                                          6.混合比例调整建议大小在0-100之间，也可以调为其他数字，但在线性组合模式下会出现未知的效果376                                          7.混合完毕后，文件将会保存在项目根目录中，文件名为output.pth377                                          8.凸组合模式会将混合比例执行Softmax使混合比例相加为1，而线性组合模式不会378                        </font>379                        """)380                    mix_model_path = gr.Files(label="选择需要混合模型文件")381                    mix_model_upload_button = gr.UploadButton("选择/追加需要混合模型文件", file_count="multiple")382                    mix_model_output1 = gr.Textbox(383                                            label="混合比例调整，单位/%",384                                            interactive = True385                                         )386                    mix_mode = gr.Radio(choices=["凸组合", "线性组合"], label="融合模式",value="凸组合",interactive = True)387                    mix_submit = gr.Button("声线融合启动", variant="primary")388                    mix_model_output2 = gr.Textbox(389                                            label="Output Message"390                                         )391                    mix_model_path.change(updata_mix_info,[mix_model_path],[mix_model_output1])392                    mix_model_upload_button.upload(upload_mix_append_file, [mix_model_upload_button,mix_model_path], [mix_model_path,mix_model_output1])393                    mix_submit.click(mix_submit_click, [mix_model_output1,mix_mode], [mix_model_output2])394                395                with gr.TabItem("模型压缩工具"):396                    gr.Markdown(value="""397                        该工具可以实现对模型的体积压缩，在**不影响模型推理功能**的情况下，将原本约600M的So-VITS模型压缩至约200M, 大大减少了硬盘的压力。398                        **注意：压缩后的模型将无法继续训练，请在确认封炉后再压缩。**399                    """)400                    model_to_compress = gr.File(label="模型上传")401                    compress_model_btn = gr.Button("压缩模型", variant="primary")402                    compress_model_output = gr.Textbox(label="输出信息", value="")403 404                    compress_model_btn.click(model_compression, [model_to_compress], [compress_model_output])405                    406                    407    with gr.Tabs():408        with gr.Row(variant="panel"):409            with gr.Column():410                gr.Markdown(value="""411                    <font size=2> WebUI设置</font>412                    """)413                debug_button = gr.Checkbox(label="Debug模式，如果向社区反馈BUG需要打开，打开后控制台可以显示具体错误提示", value=debug)414        # refresh local model list415        local_model_refresh_btn.click(local_model_refresh_fn, outputs=local_model_selection)416        # set local enabled/disabled on tab switch417        local_model_tab_upload.select(lambda: False, outputs=local_model_enabled)418        local_model_tab_local.select(lambda: True, outputs=local_model_enabled)419        420        vc_submit.click(vc_fn, [sid, vc_input3, output_format, vc_transform,auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,f0_predictor,enhancer_adaptive_key,cr_threshold,k_step,use_spk_mix,second_encoding,loudness_envelope_adjustment], [vc_output1, vc_output2])421        vc_submit2.click(vc_fn2, [text2tts, tts_lang, tts_gender, tts_rate, tts_volume, sid, output_format, vc_transform,auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,f0_predictor,enhancer_adaptive_key,cr_threshold,k_step,use_spk_mix,second_encoding,loudness_envelope_adjustment], [vc_output1, vc_output2])422 423        debug_button.change(debug_change,[],[])424        model_load_button.click(modelAnalysis,[model_path,config_path,cluster_model_path,device,enhance,diff_model_path,diff_config_path,only_diffusion,use_spk_mix,local_model_enabled,local_model_selection],[sid,sid_output])425        model_unload_button.click(modelUnload,[],[sid,sid_output])426    os.system("start http://127.0.0.1:7860")427    app.launch()428 429 430 431