recipes

This directory contains data processing scripts and training/decoding configs for performing speech recognition using wav2letter++ on popular datasets.

Preparing data

Requirements to run scripts:

• python 3
• sox and tqdm libraries, available through pip.

In general, each dataset contains prepare_data.py which prepares the Dataset and Tokens file and prepare_lm.py which prepares Lexicon and Language Model data. See README.md under each dataset directory for extra dataset-specific setups. Each file in the directory has instruction on how to run the python script.

[...]/prepare_data.py [OPTIONS ...]

[...]/prepare_lm.py [OPTIONS ...]

Training/Decoding

The configs for training and decoding can be found under configs folder. Make sure to replace [...] with appropriate paths.

To run training

[...]/wav2letter/build/Train train --flagsfile train.cfg

To run decoding

[...]/wav2letter/build/Decode --flagsfile decode.cfg

Replace [...] with appropriate paths