| 28 | |
| 29 | |
| 30 | class FlowSepPreprocessor: |
| 31 | def __init__(self, config): |
| 32 | import utilities.audio as Audio |
| 33 | |
| 34 | self.sampling_rate = config["preprocessing"]["audio"]["sampling_rate"] |
| 35 | self.duration = config["preprocessing"]["audio"]["duration"] |
| 36 | self.hopsize = config["preprocessing"]["stft"]["hop_length"] |
| 37 | self.target_length = int(self.duration * self.sampling_rate / self.hopsize) |
| 38 | |
| 39 | self.STFT = Audio.stft.TacotronSTFT( |
| 40 | config["preprocessing"]["stft"]["filter_length"], |
| 41 | config["preprocessing"]["stft"]["hop_length"], |
| 42 | config["preprocessing"]["stft"]["win_length"], |
| 43 | config["preprocessing"]["mel"]["n_mel_channels"], |
| 44 | config["preprocessing"]["audio"]["sampling_rate"], |
| 45 | config["preprocessing"]["mel"]["mel_fmin"], |
| 46 | config["preprocessing"]["mel"]["mel_fmax"], |
| 47 | ) |
| 48 | |
| 49 | def read_wav_file(self, filename): |
| 50 | waveform, sr = torchaudio.load(filename) |
| 51 | target_length = int(sr * self.duration) |
| 52 | if waveform.shape[-1] > target_length: |
| 53 | waveform = waveform[:, :target_length] |
| 54 | if sr != self.sampling_rate: |
| 55 | waveform = torchaudio.functional.resample(waveform, sr, self.sampling_rate) |
| 56 | waveform = waveform.numpy()[0, ...] |
| 57 | waveform = waveform - np.mean(waveform) |
| 58 | waveform = waveform / (np.max(np.abs(waveform)) + 1e-8) |
| 59 | waveform = waveform * 0.5 |
| 60 | waveform = waveform[None, ...] |
| 61 | target_samples = int(self.sampling_rate * self.duration) |
| 62 | if waveform.shape[-1] < target_samples: |
| 63 | temp_wav = np.zeros((1, target_samples), dtype=np.float32) |
| 64 | temp_wav[:, :waveform.shape[-1]] = waveform |
| 65 | waveform = temp_wav |
| 66 | return waveform |
| 67 | |
| 68 | def wav_feature_extraction(self, waveform): |
| 69 | import utilities.audio as Audio |
| 70 | |
| 71 | waveform = waveform[0, ...] |
| 72 | waveform = torch.FloatTensor(waveform) |
| 73 | log_mel_spec, stft, energy = Audio.tools.get_mel_from_wav(waveform, self.STFT) |
| 74 | log_mel_spec = torch.FloatTensor(log_mel_spec.T) |
| 75 | stft = torch.FloatTensor(stft.T) |
| 76 | log_mel_spec = self._pad_spec(log_mel_spec) |
| 77 | stft = self._pad_spec(stft) |
| 78 | return log_mel_spec, stft |
| 79 | |
| 80 | def _pad_spec(self, log_mel_spec): |
| 81 | n_frames = log_mel_spec.shape[0] |
| 82 | p = self.target_length - n_frames |
| 83 | if p > 0: |
| 84 | m = torch.nn.ZeroPad2d((0, 0, 0, p)) |
| 85 | log_mel_spec = m(log_mel_spec) |
| 86 | elif p < 0: |
| 87 | log_mel_spec = log_mel_spec[:self.target_length, :] |