How to speed-up feature extraction?

[desh2608]

Default feature extraction in Lhotse is taking a long time. How can I speed-up this process or perform on-the-fly feature extraction?

cc:@aarora8

[desh2608]

Answer by @pzelasko:

1. (simplest) increase the number of jobs for feature extraction
2. Use GPU to accelerate feature extraction; you will need to pip install kaldifeat for this though. See here for the relevant method + example in Lhotse: https://github.com/lhotse-speech/lhotse/blob/master/lhotse/cut.py#L3358
3. You can use on-the-fly feature extraction. It might require a bit more IO but possibly increasing num_workers in DataLoader will offset that, if you notice a slow-down in training. All the changes you need to do are here:
   https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py#L259:L267

[pzelasko]

Another viable option:

4. Return batches of waveforms from your dataset, and use lhotse.features.kaldi.layers.Wav2LogFilterBank as a layer in your neural network. Here is an example how you'd do it with K2SpeechRecognitionDataset:

from lhotse.dataset import AudioSamples, K2SpeechRecognitionDataset
dset = K2SpeechRecognitionDataset(
                ...,
                input_strategy=AudioSamples(),
            )

[csukuangfj]

Also cc @EmreOzkose (from k2-fsa/icefall#101)