Paza benchmarks and models target low resource speech recognition

Microsoft Research has introduced Paza as a human-centered speech pipeline focused on automatic speech recognition for low resource languages. Paza is designed to support speech technology development where data and tools are scarce, with an emphasis on practical usability and alignment with the needs of speakers and communities.

Alongside the pipeline, Microsoft Research is launching PazaBench, described as the first leaderboard dedicated to low-resource languages. PazaBench covers 39 African languages and 52 models and is tested with communities in real settings, providing a structured way to compare performance across a diverse set of languages and systems.

The combination of Paza and PazaBench is positioned to establish common benchmarks for low resource speech recognition and to encourage improvements in model quality for African languages. By grounding evaluations in real-world community testing, the initiative aims to make speech technologies more reliable and inclusive for underrepresented language groups.