lhotse_audio_manifest_cutset_fbank_feature_extraction.py

1import torch
2from lhotse import RecordingSet, SupervisionSet, CutSet, Fbank
3
4# 1. Create manifests from a directory of audio files
5# Assuming you have a directory 'data/audio' with .wav files
6recordings = RecordingSet.from_dir("data/audio", pattern="*.wav")
7supervisions = SupervisionSet.from_segments([]) # Or load from transcriptions
8
9# 2. Create a CutSet
10cuts = CutSet.from_manifests(recordings=recordings, supervisions=supervisions)
11
12# 3. Extract features (Filterbanks)
13# This will compute features and store them on disk
14extractor = Fbank()
15cuts_with_feats = cuts.compute_and_store_features(
16    extractor=extractor,
17    storage_path="data/feats",
18    num_jobs=4
19)
20
21# 4. Manipulation (Slicing, Padding, Concatenating)
22# Example: pad to 10 seconds
23padded_cuts = cuts_with_feats.pad(duration=10.0)
24
25# 5. Interaction with PyTorch
26# Lhotse provides datasets and samplers for easy integration
27from lhotse.dataset import KNetDataset
28from torch.utils.data import DataLoader
29
30dataset = KNetDataset()
31# Typically used with a Lhotse Sampler (e.g., CutSampler)
32# which handles bucketing by duration