README.md

Python examples

In this directory, we provide a couple of examples, described below.

Walkthrough tutorial

WavAugment_walkthrough.ipynb (open in colab) provides a succint walkthrough tutorial, showing how effects can be applied on a piece of speech (recorded over the mic or pre-recorded).

Processing a single file

The script process_file.py gives a taste on how different compositions of speech augmentation techniques sound like, by allowing to augment single files. It suppors a few randomized data augmentations: pitch, reverberation, temporal masking, band rejection, and clipping.

A typical usage is:

python process_file.py --input_file=./tests/test.wav \ 
    --output_file=augmented.wav \ 
    --chain=pitch,reverb,time_drop \ 
    --pitch_shift_max=500 \ 
    --t_ms=100

where --chain specifies a list of augmentations applied sequentially, left-to-right; t_ms and pitch_shift_max specify parameters of the augmentations. augmented.wav would contain the randomly augmented sound.

Usage in self-supervised learning

In librispeech_selfsupervised.py we use WavAugment in a way that can be used for self-supervised learning. We define a dataset that iterates over Librispeech data, reads a (randomly shifted) sequence of pre-defined length from each file and returns two copies of it, independently augmented in different ways. This example does not learns a model, only measures the dataset reading time.

The code is thoroughly documented. This command will download dev-clean in the ./data directory (if needed), iterate over it, extracting sequences of 1 second length. The batches of size 32 would be prepared by 8 DataLoader workers.

python librispeech_selfsupervised.py --data=./data --subset=dev-clean --sequence_length_seconds=1 --n_workers=8 --download --batch_size=8

Iterating over Librispeech train-clean-100 (100 hours of audio) with 16 workers takes 2 seconds without any data augmentation. With WavAugment data augmentation it takes around 5 seconds (on a solid server).