Speech, shaped closer to home

Six languages.
One listening model.

An East African speech recognition experiment that began during the AfriVoices hackathon and kept growing afterwards. Listen, record, and see how it hears language in the real world.

1 Choose what the model hears

Bring a voice

Use one of the short AfriVoices samples, upload audio, or record a few words yourself.

Transcription mode
♫
No audio selectedUp to 3 minutes · 30 MB maximum
Or begin with a sampleThree medium-length clips
2 Read what it heard

The transcript

Fast gives the model's direct reading. Accurate adds a language-aware word decoder.

Waiting for audio.
The words will appear here.
Text styleThis model writes in lowercase without punctuation. Sentence boundaries, commas and capital letters are not predicted.
Origin

Past the leaderboard

The first version was built for an ASR hackathon. Work continued after the competition, turning a submission into something people could actually try.

Purpose

Languages worth hearing

The project focuses on six languages from the region, including Kalenjin—the language that made this work personal in the first place.

Reality

An experiment, still learning

Results vary with accent, microphone and background noise. Kalenjin and Maasai remain harder for the model, and improving them is part of the journey.