Julius8888/XJP_Voice

1import torch2import torch.utils.data3from librosa.filters import mel as librosa_mel_fn4import warnings5 6# warnings.simplefilter(action='ignore', category=FutureWarning)7warnings.filterwarnings(action="ignore")8MAX_WAV_VALUE = 32768.09 10 11def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):12    """13    PARAMS14    ------15    C: compression factor16    """17    return torch.log(torch.clamp(x, min=clip_val) * C)18 19 20def dynamic_range_decompression_torch(x, C=1):21    """22    PARAMS23    ------24    C: compression factor used to compress25    """26    return torch.exp(x) / C27 28 29def spectral_normalize_torch(magnitudes):30    output = dynamic_range_compression_torch(magnitudes)31    return output32 33 34def spectral_de_normalize_torch(magnitudes):35    output = dynamic_range_decompression_torch(magnitudes)36    return output37 38 39mel_basis = {}40hann_window = {}41 42 43def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):44    if torch.min(y) < -1.0:45        print("min value is ", torch.min(y))46    if torch.max(y) > 1.0:47        print("max value is ", torch.max(y))48 49    global hann_window50    dtype_device = str(y.dtype) + "_" + str(y.device)51    wnsize_dtype_device = str(win_size) + "_" + dtype_device52    if wnsize_dtype_device not in hann_window:53        hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(54            dtype=y.dtype, device=y.device55        )56 57    y = torch.nn.functional.pad(58        y.unsqueeze(1),59        (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)),60        mode="reflect",61    )62    y = y.squeeze(1)63 64    spec = torch.stft(65        y,66        n_fft,67        hop_length=hop_size,68        win_length=win_size,69        window=hann_window[wnsize_dtype_device],70        center=center,71        pad_mode="reflect",72        normalized=False,73        onesided=True,74        return_complex=False,75    )76 77    spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)78    return spec79 80 81def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):82    global mel_basis83    dtype_device = str(spec.dtype) + "_" + str(spec.device)84    fmax_dtype_device = str(fmax) + "_" + dtype_device85    if fmax_dtype_device not in mel_basis:86        mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)87        mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(88            dtype=spec.dtype, device=spec.device89        )90    spec = torch.matmul(mel_basis[fmax_dtype_device], spec)91    spec = spectral_normalize_torch(spec)92    return spec93 94 95def mel_spectrogram_torch(96    y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False97):98    if torch.min(y) < -1.0:99        print("min value is ", torch.min(y))100    if torch.max(y) > 1.0:101        print("max value is ", torch.max(y))102 103    global mel_basis, hann_window104    dtype_device = str(y.dtype) + "_" + str(y.device)105    fmax_dtype_device = str(fmax) + "_" + dtype_device106    wnsize_dtype_device = str(win_size) + "_" + dtype_device107    if fmax_dtype_device not in mel_basis:108        mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)109        mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(110            dtype=y.dtype, device=y.device111        )112    if wnsize_dtype_device not in hann_window:113        hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(114            dtype=y.dtype, device=y.device115        )116 117    y = torch.nn.functional.pad(118        y.unsqueeze(1),119        (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)),120        mode="reflect",121    )122    y = y.squeeze(1)123 124    spec = torch.stft(125        y,126        n_fft,127        hop_length=hop_size,128        win_length=win_size,129        window=hann_window[wnsize_dtype_device],130        center=center,131        pad_mode="reflect",132        normalized=False,133        onesided=True,134        return_complex=False,135    )136 137    spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)138 139    spec = torch.matmul(mel_basis[fmax_dtype_device], spec)140    spec = spectral_normalize_torch(spec)141 142    return spec143