import torch import torchaudio import torchaudio.transforms as T from src.interfaces import IAudioProcessor from src.models.config import GrassmannConfig class AudioService(IAudioProcessor): """Handles audio loading and mel spectrogram extraction.""" def __init__(self, config: GrassmannConfig): self.config = config self.mel_transform = T.MelSpectrogram( sample_rate=config.sample_rate, n_fft=config.n_fft, hop_length=config.hop_length, n_mels=config.n_mels, ) def load(self, path: str) -> torch.Tensor: waveform, sr = torchaudio.load(path) # Resample if needed if sr != self.config.sample_rate: resampler = T.Resample(sr, self.config.sample_rate) waveform = resampler(waveform) # Convert to mono if waveform.shape[0] > 1: waveform = waveform.mean(dim=0, keepdim=True) return waveform.squeeze(0) def extract_features(self, waveform: torch.Tensor) -> torch.Tensor: if waveform.dim() == 1: waveform = waveform.unsqueeze(0) mel = self.mel_transform(waveform) # (1, n_mels, time) # Log mel mel = torch.log(mel.clamp(min=1e-9)) return mel.squeeze(0) # (n_mels, time) def normalize(self, features: torch.Tensor) -> torch.Tensor: mean = features.mean() std = features.std().clamp(min=1e-9) return (features - mean) / std