File size: 2,133 Bytes
e863e90 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 | """
===============================================================================
features/__init__.py — Feature Extraction Pipeline
===============================================================================
This package converts preprocessed audio (clean numpy arrays) into fixed-size
tensors suitable for CNN input.
Pipeline: Clean audio → Mel spectrogram → Log scale → Pad/Trim → Tensor
OWNER: sala7
===============================================================================
"""
from features.mel_spectrogram import extract_mel_spectrogram
from features.padding import pad_or_trim_spectrogram
from features.augmentation import apply_augmentation
def audio_to_tensor(audio, sr, augment=False):
"""
Convert a preprocessed audio waveform into a fixed-size tensor for CNN input.
This is the main entry point called by JSON's Dataset class. It chains
mel spectrogram extraction, padding/trimming, and optional augmentation.
Returns
-------
torch.Tensor
Shape: (1, N_MELS, FIXED_TIME_FRAMES) = (1, 128, 281)
- 1 channel (grayscale spectrogram)
- 128 mel frequency bins
- 281 time frames
Pipeline
--------
1. extract_mel_spectrogram: audio → log-mel spectrogram (128 × T)
2. pad_or_trim_spectrogram: (128 × T) → (128 × 281) fixed size
3. apply_augmentation: (optional) SpecAugment + noise injection
4. Add channel dimension: (128 × 281) → (1, 128, 281)
"""
import torch
# Step 1: Extract log-mel spectrogram
mel_spec = extract_mel_spectrogram(audio, sr)
# Step 2: Pad or trim to fixed time dimension
mel_spec = pad_or_trim_spectrogram(mel_spec)
# Step 3: Optional augmentation (training only)
if augment:
mel_spec = apply_augmentation(mel_spec)
# Step 4: Convert to tensor and add channel dimension
# Shape: (n_mels, time_frames) → (1, n_mels, time_frames)
tensor = torch.FloatTensor(mel_spec).unsqueeze(0) # this is actually important as the CNN takes three dimentions not 2
# so here we add a new empty dimension for the channel
return tensor
|