| import torch | |
| import librosa | |
| import torchaudio | |
| import numpy as np | |
| def compute_mel_spectrogram(audio, stft_processor): | |
| return stft_processor.compute_mel_spectrogram(torch.autograd.Variable(torch.clip(torch.FloatTensor(audio).unsqueeze(0), -1, 1), requires_grad=False)).squeeze(0).numpy().astype(np.float32) | |
| def pad_spectrogram(spectrogram, target_length=1024): | |
| pad_amount = target_length - spectrogram.shape[0] | |
| spectrogram = torch.nn.functional.pad(spectrogram, (0, 0, 0, pad_amount)) if pad_amount > 0 else spectrogram[:target_length, :] | |
| if spectrogram.size(-1) % 2 != 0: spectrogram = spectrogram[..., :-1] | |
| return spectrogram | |
| def pad_waveform(waveform, segment_length): | |
| waveform_length = waveform.shape[-1] | |
| assert waveform_length > 100 | |
| if segment_length is None or waveform_length == segment_length: return waveform | |
| elif waveform_length > segment_length: return waveform[:, :segment_length] | |
| padded_waveform = np.zeros((1, segment_length)) | |
| padded_waveform[:, :waveform_length] = waveform | |
| return padded_waveform | |
| def normalize(waveform): | |
| waveform -= np.mean(waveform) | |
| return (waveform / (np.max(np.abs(waveform)) + 1e-8)) * 0.5 | |
| def process_audio(y, sr, segment_length): | |
| normalized_waveform = normalize(torchaudio.functional.resample(torch.from_numpy(y), orig_freq=sr, new_freq=16000).numpy())[None, ...] | |
| return 0.5 * (pad_waveform(normalized_waveform, segment_length) / np.max(np.abs(normalized_waveform))) | |
| def load_audio(audio_path, stft_processor, device=None): | |
| y, sr = librosa.load(audio_path, sr=None) | |
| duration = len(y) / sr | |
| return pad_spectrogram(torch.FloatTensor(compute_mel_spectrogram(torch.FloatTensor(process_audio(y, sr, int(duration * 102.4) * 160)[0, ...]), stft_processor).T), int(duration * 102.4)).unsqueeze(0).unsqueeze(0).to(device), duration |