News research
February 20, 2026
Sauti Project releases benchmark for African speech recognition
TRI AI's Sauti Project has published a new benchmark for speech recognition across five under-resourced African languages, alongside baseline results from contemporary self-supervised speech models.
The Sauti Project this week released a public benchmark for automatic speech recognition (ASR) across five under-resourced African languages, with paired baseline results from contemporary self-supervised speech models.
The benchmark - and the accompanying paper, published at the African NLP Workshop - is the result of more than a year of dataset construction, evaluation design, and community review. It is intended as a reusable evaluation harness for future African ASR work, not a one-off comparison.
The full paper, evaluation code, and dataset documentation are linked from the Publications page.