Microsoft Research Launches Paza to Bring Speech Recognition to Low-Resource Languages

Microsoft Research has released Paza, a new speech recognition system designed specifically for languages that have been historically underrepresented in AI. The project includes both a benchmark (PazaBench) and fine-tuned models covering 39 African languages.

The name comes from the Swahili phrase "paza sauti," meaning "to project" or "to raise your voice"—a deliberate choice reflecting Microsoft's stated goal of co-creating speech technologies with the communities they aim to serve.

"We found that speech systems often fail in real-world, low-resource environments—where many languages go unrecognized and non-Western accents are frequently misunderstood," Microsoft noted. The project grew out of Project Gecko, a collaboration with Digital Green that built AI tools for farmers across Africa and India.

PazaBench is being called the first ASR leaderboard for low-resource languages, tracking 51 models across 39 African languages using three metrics: Character Error Rate, Word Error Rate, and RTFx (inverse real-time factor, measuring transcription speed).

The Paza ASR models cover six Kenyan languages: Swahili, Dholuo, Kalenjin, Kikuyu, Maasai, and Somali. Early versions were tested directly with farmers on everyday mobile devices, with feedback driving subsequent fine-tuning.

"Rather than simply adding more languages to existing systems, Paza is about co-creating speech technologies in partnership with the communities who use them," Microsoft said.