Building an Azerbaijani speech recognition model with 212 people

Speech recognition works well for English because there is an enormous amount of English audio to train on. Azerbaijani does not have that luxury. To build a good model, I first had to build the dataset.

Collecting the data

I ran a two month crowdsourcing competition to collect Azerbaijani speech. 212 participants contributed recordings, and by the end we had 750 hours of speech to train on.

A competition turned data collection from a chore into something people wanted to take part in, and it brought in a wide range of voices instead of a handful of speakers.

Where it is used

The resulting speech to text model is used by ASAN, the State Agency for Public Service and Social Innovations under the President of Azerbaijan, for customer service.

Why it matters

Languages with less available data are easy for large tech companies to deprioritize. This project showed that a focused effort with a motivated community can close a lot of that gap.