Voice typing for African languages

For many people, using your voice to dictate text messages, emails and other text-based communications is easier and more accessible than typing on a keyboard. Voice typing (i.e. using your voice to dictate text) is currently not available for most of the 2,000+ languages spoken in Africa. In recent years, more data for African languages has become widely available, paving the way for this technology to become a reality for many more languages. As part of our commitment to serve Africa and its languages, we are excited to announce the release of voice typing for 9 more African languages.
  • isiNdebele
  • isiXhosa
  • Kinyarwanda
  • Northern Sotho
  • Swati
  • Sesotho
  • Tswana
  • Tshivenda
  • Xitsonga

Along with the four African languages we already support: Afrikaans, Amharic, Swahili and Zulu, this release brings voice typing support to 13 African languages, and 80 languages total around the world.

This development would not have been possible without two key advances in the state of the art for automatic speech recognition. The first advance is related to AI models for speech recognition. We employed a technique known as multilingual modeling, which uses data from multiple languages to train a single speech recognition model. This method allows the languages with less data to benefit from those with more data, to improve quality for all the languages. The second advance is related to data. In the last few years, communities, individuals and organizations have created and open sourced high quality datasets for African languages. 

The languages launching today are possible thanks in part to the efforts of researchers and organizations in Africa to create and publish data (see our paper for the data we used for each language). In particular, we’d like to thank the creators of the NCHLT corpus for South African languages, without which many of the South African languages launching today would not have been possible. We’d also like to thank Digital Umuganda for their work in creating the Kinyarwanda corpus and publishing it on Mozilla Common Voice, one of the largest resources ever created for an African language. 

Google is also working to collect data for more African languages, through our TaskMate and Crowdsource platforms, and we have partnered with universities and researchers on data collection projects, for example our work with the Bambara community and the Waxal speech data project.

Wherever you want to type, whether it’s a message, an email, or posting on social media, try voice typing on your Android device with Gboard. It’s quick, easy and faster for your friends to read than a voice note :)

Get set up on Gboard: 

Posted by Sandy Ritchie, Linguist in the Speech Recognition team