mosibi/RVC_HFv2

1import numpy as np2 3 4# This function is obtained from librosa.5def get_rms(6    y,7    frame_length=2048,8    hop_length=512,9    pad_mode="constant",10):11    padding = (int(frame_length // 2), int(frame_length // 2))12    y = np.pad(y, padding, mode=pad_mode)13 14    axis = -115    # put our new within-frame axis at the end for now16    out_strides = y.strides + tuple([y.strides[axis]])17    # Reduce the shape on the framing axis18    x_shape_trimmed = list(y.shape)19    x_shape_trimmed[axis] -= frame_length - 120    out_shape = tuple(x_shape_trimmed) + tuple([frame_length])21    xw = np.lib.stride_tricks.as_strided(y, shape=out_shape, strides=out_strides)22    if axis < 0:23        target_axis = axis - 124    else:25        target_axis = axis + 126    xw = np.moveaxis(xw, -1, target_axis)27    # Downsample along the target axis28    slices = [slice(None)] * xw.ndim29    slices[axis] = slice(0, None, hop_length)30    x = xw[tuple(slices)]31 32    # Calculate power33    power = np.mean(np.abs(x) ** 2, axis=-2, keepdims=True)34 35    return np.sqrt(power)36 37 38class Slicer:39    def __init__(40        self,41        sr: int,42        threshold: float = -40.0,43        min_length: int = 5000,44        min_interval: int = 300,45        hop_size: int = 20,46        max_sil_kept: int = 5000,47    ):48        if not min_length >= min_interval >= hop_size:49            raise ValueError(50                "The following condition must be satisfied: min_length >= min_interval >= hop_size"51            )52        if not max_sil_kept >= hop_size:53            raise ValueError(54                "The following condition must be satisfied: max_sil_kept >= hop_size"55            )56        min_interval = sr * min_interval / 100057        self.threshold = 10 ** (threshold / 20.0)58        self.hop_size = round(sr * hop_size / 1000)59        self.win_size = min(round(min_interval), 4 * self.hop_size)60        self.min_length = round(sr * min_length / 1000 / self.hop_size)61        self.min_interval = round(min_interval / self.hop_size)62        self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)63 64    def _apply_slice(self, waveform, begin, end):65        if len(waveform.shape) > 1:66            return waveform[67                :, begin * self.hop_size : min(waveform.shape[1], end * self.hop_size)68            ]69        else:70            return waveform[71                begin * self.hop_size : min(waveform.shape[0], end * self.hop_size)72            ]73 74    # @timeit75    def slice(self, waveform):76        if len(waveform.shape) > 1:77            samples = waveform.mean(axis=0)78        else:79            samples = waveform80        if samples.shape[0] <= self.min_length:81            return [waveform]82        rms_list = get_rms(83            y=samples, frame_length=self.win_size, hop_length=self.hop_size84        ).squeeze(0)85        sil_tags = []86        silence_start = None87        clip_start = 088        for i, rms in enumerate(rms_list):89            # Keep looping while frame is silent.90            if rms < self.threshold:91                # Record start of silent frames.92                if silence_start is None:93                    silence_start = i94                continue95            # Keep looping while frame is not silent and silence start has not been recorded.96            if silence_start is None:97                continue98            # Clear recorded silence start if interval is not enough or clip is too short99            is_leading_silence = silence_start == 0 and i > self.max_sil_kept100            need_slice_middle = (101                i - silence_start >= self.min_interval102                and i - clip_start >= self.min_length103            )104            if not is_leading_silence and not need_slice_middle:105                silence_start = None106                continue107            # Need slicing. Record the range of silent frames to be removed.108            if i - silence_start <= self.max_sil_kept:109                pos = rms_list[silence_start : i + 1].argmin() + silence_start110                if silence_start == 0:111                    sil_tags.append((0, pos))112                else:113                    sil_tags.append((pos, pos))114                clip_start = pos115            elif i - silence_start <= self.max_sil_kept * 2:116                pos = rms_list[117                    i - self.max_sil_kept : silence_start + self.max_sil_kept + 1118                ].argmin()119                pos += i - self.max_sil_kept120                pos_l = (121                    rms_list[122                        silence_start : silence_start + self.max_sil_kept + 1123                    ].argmin()124                    + silence_start125                )126                pos_r = (127                    rms_list[i - self.max_sil_kept : i + 1].argmin()128                    + i129                    - self.max_sil_kept130                )131                if silence_start == 0:132                    sil_tags.append((0, pos_r))133                    clip_start = pos_r134                else:135                    sil_tags.append((min(pos_l, pos), max(pos_r, pos)))136                    clip_start = max(pos_r, pos)137            else:138                pos_l = (139                    rms_list[140                        silence_start : silence_start + self.max_sil_kept + 1141                    ].argmin()142                    + silence_start143                )144                pos_r = (145                    rms_list[i - self.max_sil_kept : i + 1].argmin()146                    + i147                    - self.max_sil_kept148                )149                if silence_start == 0:150                    sil_tags.append((0, pos_r))151                else:152                    sil_tags.append((pos_l, pos_r))153                clip_start = pos_r154            silence_start = None155        # Deal with trailing silence.156        total_frames = rms_list.shape[0]157        if (158            silence_start is not None159            and total_frames - silence_start >= self.min_interval160        ):161            silence_end = min(total_frames, silence_start + self.max_sil_kept)162            pos = rms_list[silence_start : silence_end + 1].argmin() + silence_start163            sil_tags.append((pos, total_frames + 1))164        # Apply and return slices.165        if len(sil_tags) == 0:166            return [waveform]167        else:168            chunks = []169            if sil_tags[0][0] > 0:170                chunks.append(self._apply_slice(waveform, 0, sil_tags[0][0]))171            for i in range(len(sil_tags) - 1):172                chunks.append(173                    self._apply_slice(waveform, sil_tags[i][1], sil_tags[i + 1][0])174                )175            if sil_tags[-1][1] < total_frames:176                chunks.append(177                    self._apply_slice(waveform, sil_tags[-1][1], total_frames)178                )179            return chunks180 181 182def main():183    import os.path184    from argparse import ArgumentParser185 186    import librosa187    import soundfile188 189    parser = ArgumentParser()190    parser.add_argument("audio", type=str, help="The audio to be sliced")191    parser.add_argument(192        "--out", type=str, help="Output directory of the sliced audio clips"193    )194    parser.add_argument(195        "--db_thresh",196        type=float,197        required=False,198        default=-40,199        help="The dB threshold for silence detection",200    )201    parser.add_argument(202        "--min_length",203        type=int,204        required=False,205        default=5000,206        help="The minimum milliseconds required for each sliced audio clip",207    )208    parser.add_argument(209        "--min_interval",210        type=int,211        required=False,212        default=300,213        help="The minimum milliseconds for a silence part to be sliced",214    )215    parser.add_argument(216        "--hop_size",217        type=int,218        required=False,219        default=10,220        help="Frame length in milliseconds",221    )222    parser.add_argument(223        "--max_sil_kept",224        type=int,225        required=False,226        default=500,227        help="The maximum silence length kept around the sliced clip, presented in milliseconds",228    )229    args = parser.parse_args()230    out = args.out231    if out is None:232        out = os.path.dirname(os.path.abspath(args.audio))233    audio, sr = librosa.load(args.audio, sr=None, mono=False)234    slicer = Slicer(235        sr=sr,236        threshold=args.db_thresh,237        min_length=args.min_length,238        min_interval=args.min_interval,239        hop_size=args.hop_size,240        max_sil_kept=args.max_sil_kept,241    )242    chunks = slicer.slice(audio)243    if not os.path.exists(out):244        os.makedirs(out)245    for i, chunk in enumerate(chunks):246        if len(chunk.shape) > 1:247            chunk = chunk.T248        soundfile.write(249            os.path.join(250                out,251                f"%s_%d.wav"252                % (os.path.basename(args.audio).rsplit(".", maxsplit=1)[0], i),253            ),254            chunk,255            sr,256        )257 258 259if __name__ == "__main__":260    main()261