Microsoft Research has launched Paza, a project aimed at improving speech recognition for African languages that are often ignored by mainstream AI systems. The initiative targets a clear gap: as voice tools become a common way to access digital services, many languages spoken across the continent still lack adequate support.
Paza combines two parts.
The first is PazaBench, a benchmarking platform for testing speech recognition systems in low-resource languages. It covers 39 African languages and evaluates 52 existing speech and language models. It also includes new speech models released for six Kenyan languages.
The second is a set of speech recognition models fine-tuned for everyday use in Swahili, Dholuo, Kalenjin, Kikuyu, Maasai, and Somali. The models were tested in noisy, low-bandwidth environments to reflect real-world usage rather than lab conditions.
Microsoft researchers say the focus is on practical performance. The system aims to work on standard mobile devices and for users who rely on spoken interaction rather than written text.
The project also relies on community involvement during data collection and testing, a response to criticism that many AI tools fail when faced with accents, dialects, or local contexts.
PazaBench is positioned as an open platform that allows researchers and developers to compare model performance and identify where systems fail. Microsoft says this transparency should accelerate improvements and reduce duplication across research teams.
The company plans to expand the benchmark to more languages and publish guidance on dataset building and model tuning, hoping to encourage broader participation from universities, startups, and local developers.
Paza takes its name from a Swahili word meaning “to raise your voice,” reflecting the project’s focus on giving more languages a place in speech technology as AI systems become more central to communication and access.