| """16 kHz log-mel, exactly as Swar's style encoders were trained on it.""" | |
| from __future__ import annotations | |
| import torch | |
| import torchaudio | |
| from torch import nn | |
| SR, N_FFT, WIN, HOP, N_MELS = 16000, 1024, 800, 200, 80 | |
| MEAN, STD = -4.0, 4.0 | |
| class Mel(nn.Module): | |
| def __init__(self): | |
| super().__init__() | |
| self.mel = torchaudio.transforms.MelSpectrogram( | |
| sample_rate=SR, n_fft=N_FFT, win_length=WIN, hop_length=HOP, n_mels=N_MELS, | |
| f_min=0.0, f_max=8000.0, center=True, pad_mode="reflect", power=2.0) | |
| def forward(self, wav: torch.Tensor) -> torch.Tensor: | |
| if wav.dtype == torch.int16: | |
| wav = wav.float() / 32768.0 | |
| if wav.dim() == 1: | |
| wav = wav[None] | |
| return (torch.log(1e-5 + self.mel(wav)) - MEAN) / STD | |
| def frames(n_samples: int) -> int: | |
| return n_samples // HOP + 1 | |