mosi77/5

1import torch
2import torch.utils.data
3from librosa.filters import mel as librosa_mel_fn
4
5
6def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
7    return torch.log(torch.clamp(x, min=clip_val) * C)
8
9
10def dynamic_range_decompression_torch(x, C=1):
11    return torch.exp(x) / C
12
13
14def spectral_normalize_torch(magnitudes):
15    return dynamic_range_compression_torch(magnitudes)
16
17
18def spectral_de_normalize_torch(magnitudes):
19    return dynamic_range_decompression_torch(magnitudes)
20
21
22mel_basis = {}
23hann_window = {}
24
25
26def spectrogram_torch(y, n_fft, hop_size, win_size, center=False):
27    global hann_window
28    dtype_device = str(y.dtype) + "_" + str(y.device)
29    wnsize_dtype_device = str(win_size) + "_" + dtype_device
30    if wnsize_dtype_device not in hann_window:
31        hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(
32            dtype=y.dtype, device=y.device
33        )
34
35    y = torch.nn.functional.pad(
36        y.unsqueeze(1),
37        (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)),
38        mode="reflect",
39    )
40    y = y.squeeze(1)
41
42    spec = torch.stft(
43        y,
44        n_fft,
45        hop_length=hop_size,
46        win_length=win_size,
47        window=hann_window[wnsize_dtype_device],
48        center=center,
49        pad_mode="reflect",
50        normalized=False,
51        onesided=True,
52        return_complex=True,
53    )
54
55    spec = torch.sqrt(spec.real.pow(2) + spec.imag.pow(2) + 1e-6)
56
57    return spec
58
59
60def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
61    global mel_basis
62    dtype_device = str(spec.dtype) + "_" + str(spec.device)
63    fmax_dtype_device = str(fmax) + "_" + dtype_device
64    if fmax_dtype_device not in mel_basis:
65        mel = librosa_mel_fn(
66            sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax
67        )
68        mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(
69            dtype=spec.dtype, device=spec.device
70        )
71
72    melspec = torch.matmul(mel_basis[fmax_dtype_device], spec)
73    melspec = spectral_normalize_torch(melspec)
74    return melspec
75
76
77def mel_spectrogram_torch(
78    y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False
79):
80    spec = spectrogram_torch(y, n_fft, hop_size, win_size, center)
81
82    melspec = spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax)
83
84    return melspec
85