ChazzyG/Retrieval-based-Voice-Conversion-WebUI

1import os, sys2 3now_dir = os.getcwd()4sys.path.append(now_dir)5import PySimpleGUI as sg6import sounddevice as sd7import noisereduce as nr8import numpy as np9from fairseq import checkpoint_utils10import librosa, torch, pyworld, faiss, time, threading11import torch.nn.functional as F12import torchaudio.transforms as tat13import scipy.signal as signal14 15# import matplotlib.pyplot as plt16from infer_pack.models import SynthesizerTrnMs256NSFsid, SynthesizerTrnMs256NSFsid_nono17from i18n import I18nAuto18 19i18n = I18nAuto()20device = torch.device("cuda" if torch.cuda.is_available() else "cpu")21 22 23class RVC:24    def __init__(25        self, key, hubert_path, pth_path, index_path, npy_path, index_rate26    ) -> None:27        """28        初始化29        """30        try:31            self.f0_up_key = key32            self.time_step = 160 / 16000 * 100033            self.f0_min = 5034            self.f0_max = 110035            self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)36            self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)37            self.sr = 1600038            self.window = 16039            if index_rate != 0:40                self.index = faiss.read_index(index_path)41                # self.big_npy = np.load(npy_path)42                self.big_npy = index.reconstruct_n(0, self.index.ntotal)43                print("index search enabled")44            self.index_rate = index_rate45            model_path = hubert_path46            print("load model(s) from {}".format(model_path))47            models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(48                [model_path],49                suffix="",50            )51            self.model = models[0]52            self.model = self.model.to(device)53            self.model = self.model.half()54            self.model.eval()55            cpt = torch.load(pth_path, map_location="cpu")56            self.tgt_sr = cpt["config"][-1]57            cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]  # n_spk58            self.if_f0 = cpt.get("f0", 1)59            if self.if_f0 == 1:60                self.net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=True)61            else:62                self.net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])63            del self.net_g.enc_q64            print(self.net_g.load_state_dict(cpt["weight"], strict=False))65            self.net_g.eval().to(device)66            self.net_g.half()67        except Exception as e:68            print(e)69 70    def get_f0(self, x, f0_up_key, inp_f0=None):71        x_pad = 172        f0_min = 5073        f0_max = 110074        f0_mel_min = 1127 * np.log(1 + f0_min / 700)75        f0_mel_max = 1127 * np.log(1 + f0_max / 700)76        f0, t = pyworld.harvest(77            x.astype(np.double),78            fs=self.sr,79            f0_ceil=f0_max,80            f0_floor=f0_min,81            frame_period=10,82        )83        f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.sr)84        f0 = signal.medfilt(f0, 3)85        f0 *= pow(2, f0_up_key / 12)86        # with open("test.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))87        tf0 = self.sr // self.window  # 每秒f0点数88        if inp_f0 is not None:89            delta_t = np.round(90                (inp_f0[:, 0].max() - inp_f0[:, 0].min()) * tf0 + 191            ).astype("int16")92            replace_f0 = np.interp(93                list(range(delta_t)), inp_f0[:, 0] * 100, inp_f0[:, 1]94            )95            shape = f0[x_pad * tf0 : x_pad * tf0 + len(replace_f0)].shape[0]96            f0[x_pad * tf0 : x_pad * tf0 + len(replace_f0)] = replace_f0[:shape]97        # with open("test_opt.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))98        f0bak = f0.copy()99        f0_mel = 1127 * np.log(1 + f0 / 700)100        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (101            f0_mel_max - f0_mel_min102        ) + 1103        f0_mel[f0_mel <= 1] = 1104        f0_mel[f0_mel > 255] = 255105        f0_coarse = np.rint(f0_mel).astype(np.int)106        return f0_coarse, f0bak  # 1-0107 108    def infer(self, feats: torch.Tensor) -> np.ndarray:109        """110        推理函数111        """112        audio = feats.clone().cpu().numpy()113        assert feats.dim() == 1, feats.dim()114        feats = feats.view(1, -1)115        padding_mask = torch.BoolTensor(feats.shape).fill_(False)116        inputs = {117            "source": feats.half().to(device),118            "padding_mask": padding_mask.to(device),119            "output_layer": 9,  # layer 9120        }121        torch.cuda.synchronize()122        with torch.no_grad():123            logits = self.model.extract_features(**inputs)124            feats = self.model.final_proj(logits[0])125 126        ####索引优化127        if hasattr(self, "index") and hasattr(self, "big_npy") and self.index_rate != 0:128            npy = feats[0].cpu().numpy().astype("float32")129 130            # _, I = self.index.search(npy, 1)131            # npy = self.big_npy[I.squeeze()].astype("float16")132 133            score, ix = index.search(npy, k=8)134            weight = np.square(1 / score)135            weight /= weight.sum(axis=1, keepdims=True)136            npy = np.sum(big_npy[ix] * np.expand_dims(weight, axis=2), axis=1).astype(137                "float16"138            )139 140            feats = (141                torch.from_numpy(npy).unsqueeze(0).to(device) * self.index_rate142                + (1 - self.index_rate) * feats143            )144        else:145            print("index search FAIL or disabled")146 147        feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)148        torch.cuda.synchronize()149        print(feats.shape)150        if self.if_f0 == 1:151            pitch, pitchf = self.get_f0(audio, self.f0_up_key)152            p_len = min(feats.shape[1], 13000, pitch.shape[0])  # 太大了爆显存153        else:154            pitch, pitchf = None, None155            p_len = min(feats.shape[1], 13000)  # 太大了爆显存156        torch.cuda.synchronize()157        # print(feats.shape,pitch.shape)158        feats = feats[:, :p_len, :]159        if self.if_f0 == 1:160            pitch = pitch[:p_len]161            pitchf = pitchf[:p_len]162            pitch = torch.LongTensor(pitch).unsqueeze(0).to(device)163            pitchf = torch.FloatTensor(pitchf).unsqueeze(0).to(device)164        p_len = torch.LongTensor([p_len]).to(device)165        ii = 0  # sid166        sid = torch.LongTensor([ii]).to(device)167        with torch.no_grad():168            if self.if_f0 == 1:169                infered_audio = (170                    self.net_g.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0]171                    .data.cpu()172                    .float()173                )174            else:175                infered_audio = (176                    self.net_g.infer(feats, p_len, sid)[0][0, 0].data.cpu().float()177                )178        torch.cuda.synchronize()179        return infered_audio180 181 182class Config:183    def __init__(self) -> None:184        self.hubert_path: str = ""185        self.pth_path: str = ""186        self.index_path: str = ""187        self.npy_path: str = ""188        self.pitch: int = 12189        self.samplerate: int = 44100190        self.block_time: float = 1.0  # s191        self.buffer_num: int = 1192        self.threhold: int = -30193        self.crossfade_time: float = 0.08194        self.extra_time: float = 0.04195        self.I_noise_reduce = False196        self.O_noise_reduce = False197        self.index_rate = 0.3198 199 200class GUI:201    def __init__(self) -> None:202        self.config = Config()203        self.flag_vc = False204 205        self.launcher()206 207    def launcher(self):208        sg.theme("LightBlue3")209        input_devices, output_devices, _, _ = self.get_devices()210        layout = [211            [212                sg.Frame(213                    title=i18n("加载模型"),214                    layout=[215                        [216                            sg.Input(default_text="hubert_base.pt", key="hubert_path"),217                            sg.FileBrowse(i18n("Hubert模型")),218                        ],219                        [220                            sg.Input(default_text="TEMP\\atri.pth", key="pth_path"),221                            sg.FileBrowse(i18n("选择.pth文件")),222                        ],223                        [224                            sg.Input(225                                default_text="TEMP\\added_IVF512_Flat_atri_baseline_src_feat.index",226                                key="index_path",227                            ),228                            sg.FileBrowse(i18n("选择.index文件")),229                        ],230                        [231                            sg.Input(232                                default_text="你不需要填写这个You don't need write this.",233                                key="npy_path",234                            ),235                            sg.FileBrowse(i18n("选择.npy文件")),236                        ],237                    ],238                )239            ],240            [241                sg.Frame(242                    layout=[243                        [244                            sg.Text(i18n("输入设备")),245                            sg.Combo(246                                input_devices,247                                key="sg_input_device",248                                default_value=input_devices[sd.default.device[0]],249                            ),250                        ],251                        [252                            sg.Text(i18n("输出设备")),253                            sg.Combo(254                                output_devices,255                                key="sg_output_device",256                                default_value=output_devices[sd.default.device[1]],257                            ),258                        ],259                    ],260                    title=i18n("音频设备(请使用同种类驱动)"),261                )262            ],263            [264                sg.Frame(265                    layout=[266                        [267                            sg.Text(i18n("响应阈值")),268                            sg.Slider(269                                range=(-60, 0),270                                key="threhold",271                                resolution=1,272                                orientation="h",273                                default_value=-30,274                            ),275                        ],276                        [277                            sg.Text(i18n("音调设置")),278                            sg.Slider(279                                range=(-24, 24),280                                key="pitch",281                                resolution=1,282                                orientation="h",283                                default_value=12,284                            ),285                        ],286                        [287                            sg.Text(i18n("Index Rate")),288                            sg.Slider(289                                range=(0.0, 1.0),290                                key="index_rate",291                                resolution=0.01,292                                orientation="h",293                                default_value=0.5,294                            ),295                        ],296                    ],297                    title=i18n("常规设置"),298                ),299                sg.Frame(300                    layout=[301                        [302                            sg.Text(i18n("采样长度")),303                            sg.Slider(304                                range=(0.1, 3.0),305                                key="block_time",306                                resolution=0.1,307                                orientation="h",308                                default_value=1.0,309                            ),310                        ],311                        [312                            sg.Text(i18n("淡入淡出长度")),313                            sg.Slider(314                                range=(0.01, 0.15),315                                key="crossfade_length",316                                resolution=0.01,317                                orientation="h",318                                default_value=0.08,319                            ),320                        ],321                        [322                            sg.Text(i18n("额外推理时长")),323                            sg.Slider(324                                range=(0.05, 3.00),325                                key="extra_time",326                                resolution=0.01,327                                orientation="h",328                                default_value=0.05,329                            ),330                        ],331                        [332                            sg.Checkbox(i18n("输入降噪"), key="I_noise_reduce"),333                            sg.Checkbox(i18n("输出降噪"), key="O_noise_reduce"),334                        ],335                    ],336                    title=i18n("性能设置"),337                ),338            ],339            [340                sg.Button(i18n("开始音频转换"), key="start_vc"),341                sg.Button(i18n("停止音频转换"), key="stop_vc"),342                sg.Text(i18n("推理时间(ms):")),343                sg.Text("0", key="infer_time"),344            ],345        ]346 347        self.window = sg.Window("RVC - GUI", layout=layout)348        self.event_handler()349 350    def event_handler(self):351        while True:352            event, values = self.window.read()353            if event == sg.WINDOW_CLOSED:354                self.flag_vc = False355                exit()356            if event == "start_vc" and self.flag_vc == False:357                self.set_values(values)358                print(str(self.config.__dict__))359                print("using_cuda:" + str(torch.cuda.is_available()))360                self.start_vc()361            if event == "stop_vc" and self.flag_vc == True:362                self.flag_vc = False363 364    def set_values(self, values):365        self.set_devices(values["sg_input_device"], values["sg_output_device"])366        self.config.hubert_path = values["hubert_path"]367        self.config.pth_path = values["pth_path"]368        self.config.index_path = values["index_path"]369        self.config.npy_path = values["npy_path"]370        self.config.threhold = values["threhold"]371        self.config.pitch = values["pitch"]372        self.config.block_time = values["block_time"]373        self.config.crossfade_time = values["crossfade_length"]374        self.config.extra_time = values["extra_time"]375        self.config.I_noise_reduce = values["I_noise_reduce"]376        self.config.O_noise_reduce = values["O_noise_reduce"]377        self.config.index_rate = values["index_rate"]378 379    def start_vc(self):380        torch.cuda.empty_cache()381        self.flag_vc = True382        self.block_frame = int(self.config.block_time * self.config.samplerate)383        self.crossfade_frame = int(self.config.crossfade_time * self.config.samplerate)384        self.sola_search_frame = int(0.012 * self.config.samplerate)385        self.delay_frame = int(0.01 * self.config.samplerate)  # 往前预留0.02s386        self.extra_frame = int(self.config.extra_time * self.config.samplerate)387        self.rvc = None388        self.rvc = RVC(389            self.config.pitch,390            self.config.hubert_path,391            self.config.pth_path,392            self.config.index_path,393            self.config.npy_path,394            self.config.index_rate,395        )396        self.input_wav: np.ndarray = np.zeros(397            self.extra_frame398            + self.crossfade_frame399            + self.sola_search_frame400            + self.block_frame,401            dtype="float32",402        )403        self.output_wav: torch.Tensor = torch.zeros(404            self.block_frame, device=device, dtype=torch.float32405        )406        self.sola_buffer: torch.Tensor = torch.zeros(407            self.crossfade_frame, device=device, dtype=torch.float32408        )409        self.fade_in_window: torch.Tensor = torch.linspace(410            0.0, 1.0, steps=self.crossfade_frame, device=device, dtype=torch.float32411        )412        self.fade_out_window: torch.Tensor = 1 - self.fade_in_window413        self.resampler1 = tat.Resample(414            orig_freq=self.config.samplerate, new_freq=16000, dtype=torch.float32415        )416        self.resampler2 = tat.Resample(417            orig_freq=self.rvc.tgt_sr,418            new_freq=self.config.samplerate,419            dtype=torch.float32,420        )421        thread_vc = threading.Thread(target=self.soundinput)422        thread_vc.start()423 424    def soundinput(self):425        """426        接受音频输入427        """428        with sd.Stream(429            callback=self.audio_callback,430            blocksize=self.block_frame,431            samplerate=self.config.samplerate,432            dtype="float32",433        ):434            while self.flag_vc:435                time.sleep(self.config.block_time)436                print("Audio block passed.")437        print("ENDing VC")438 439    def audio_callback(440        self, indata: np.ndarray, outdata: np.ndarray, frames, times, status441    ):442        """443        音频处理444        """445        start_time = time.perf_counter()446        indata = librosa.to_mono(indata.T)447        if self.config.I_noise_reduce:448            indata[:] = nr.reduce_noise(y=indata, sr=self.config.samplerate)449 450        """noise gate"""451        frame_length = 2048452        hop_length = 1024453        rms = librosa.feature.rms(454            y=indata, frame_length=frame_length, hop_length=hop_length455        )456        db_threhold = librosa.amplitude_to_db(rms, ref=1.0)[0] < self.config.threhold457        # print(rms.shape,db.shape,db)458        for i in range(db_threhold.shape[0]):459            if db_threhold[i]:460                indata[i * hop_length : (i + 1) * hop_length] = 0461        self.input_wav[:] = np.append(self.input_wav[self.block_frame :], indata)462 463        # infer464        print("input_wav:" + str(self.input_wav.shape))465        # print('infered_wav:'+str(infer_wav.shape))466        infer_wav: torch.Tensor = self.resampler2(467            self.rvc.infer(self.resampler1(torch.from_numpy(self.input_wav)))468        )[-self.crossfade_frame - self.sola_search_frame - self.block_frame :].to(469            device470        )471        print("infer_wav:" + str(infer_wav.shape))472 473        # SOLA algorithm from https://github.com/yxlllc/DDSP-SVC474        cor_nom = F.conv1d(475            infer_wav[None, None, : self.crossfade_frame + self.sola_search_frame],476            self.sola_buffer[None, None, :],477        )478        cor_den = torch.sqrt(479            F.conv1d(480                infer_wav[None, None, : self.crossfade_frame + self.sola_search_frame]481                ** 2,482                torch.ones(1, 1, self.crossfade_frame, device=device),483            )484            + 1e-8485        )486        sola_offset = torch.argmax(cor_nom[0, 0] / cor_den[0, 0])487        print("sola offset: " + str(int(sola_offset)))488 489        # crossfade490        self.output_wav[:] = infer_wav[sola_offset : sola_offset + self.block_frame]491        self.output_wav[: self.crossfade_frame] *= self.fade_in_window492        self.output_wav[: self.crossfade_frame] += self.sola_buffer[:]493        if sola_offset < self.sola_search_frame:494            self.sola_buffer[:] = (495                infer_wav[496                    -self.sola_search_frame497                    - self.crossfade_frame498                    + sola_offset : -self.sola_search_frame499                    + sola_offset500                ]501                * self.fade_out_window502            )503        else:504            self.sola_buffer[:] = (505                infer_wav[-self.crossfade_frame :] * self.fade_out_window506            )507 508        if self.config.O_noise_reduce:509            outdata[:] = np.tile(510                nr.reduce_noise(511                    y=self.output_wav[:].cpu().numpy(), sr=self.config.samplerate512                ),513                (2, 1),514            ).T515        else:516            outdata[:] = self.output_wav[:].repeat(2, 1).t().cpu().numpy()517        total_time = time.perf_counter() - start_time518        self.window["infer_time"].update(int(total_time * 1000))519        print("infer time:" + str(total_time))520 521    def get_devices(self, update: bool = True):522        """获取设备列表"""523        if update:524            sd._terminate()525            sd._initialize()526        devices = sd.query_devices()527        hostapis = sd.query_hostapis()528        for hostapi in hostapis:529            for device_idx in hostapi["devices"]:530                devices[device_idx]["hostapi_name"] = hostapi["name"]531        input_devices = [532            f"{d['name']} ({d['hostapi_name']})"533            for d in devices534            if d["max_input_channels"] > 0535        ]536        output_devices = [537            f"{d['name']} ({d['hostapi_name']})"538            for d in devices539            if d["max_output_channels"] > 0540        ]541        input_devices_indices = [542            d["index"] for d in devices if d["max_input_channels"] > 0543        ]544        output_devices_indices = [545            d["index"] for d in devices if d["max_output_channels"] > 0546        ]547        return (548            input_devices,549            output_devices,550            input_devices_indices,551            output_devices_indices,552        )553 554    def set_devices(self, input_device, output_device):555        """设置输出设备"""556        (557            input_devices,558            output_devices,559            input_device_indices,560            output_device_indices,561        ) = self.get_devices()562        sd.default.device[0] = input_device_indices[input_devices.index(input_device)]563        sd.default.device[1] = output_device_indices[564            output_devices.index(output_device)565        ]566        print("input device:" + str(sd.default.device[0]) + ":" + str(input_device))567        print("output device:" + str(sd.default.device[1]) + ":" + str(output_device))568 569 570gui = GUI()571