Start where we stand
We begin with Gĩkũyũ because we speak it. We can hear when the model gets a tone wrong, judge a transcript like a native, and build with the community it serves instead of from a distance.
Aria, twandĩke.
Our languages deserve to be heard and written.
Afrika Labs is building open speech-to-text for African languages, starting with Gĩkũyũ. Speak naturally and watch your words become text.
"Aria, twandĩke" is Gĩkũyũ for "Speak, let us write."
Mũgambo · Voice
Speak Gĩkũyũ. We're learning to listen.
Press the microphone, say something in Gĩkũyũ, and hear it back. Transcription switches on the moment our first model finishes training.
Your recording
Audio stays in your browser. Nothing is uploaded until a transcription endpoint is connected — and we'll always say so here first.
Rũthiomi · Language
Speech technology skipped our languages.
Voice assistants, dictation, and subtitles exist because someone trained a model. For many African languages, the data and models are still limited. Afrika Labs is working on that gap, one language at a time.
8M+
people speak Gĩkũyũ today
2,000+
languages are spoken across Africa
217h
of transcribed Gĩkũyũ speech in our first training corpus
Built in the open
We publish our progress, benchmarks, and mistakes. Speech technology for African languages should be something the continent can inspect, improve, and own.
A method, not just a model
Gathering voices, training models, and evaluating with native speakers should form a repeatable pipeline. Gĩkũyũ is where we start. Kiswahili, Dholuo, and other languages can follow.
Rũgendo · The journey
Build the speech stack one useful release at a time.
-
Now
Gĩkũyũ v0
Train and evaluate the first Gĩkũyũ speech-to-text model, publish the benchmark and model, and make it available through a public demo. This version is experimental and focused on proving the pipeline end to end.
-
Next
Gĩkũyũ v1
Turn the first model into production-grade ASR with better accuracy, broader native-speaker evaluation, stronger performance across dialects and noisy audio, faster inference, long-form and streaming support, and stable documentation and APIs.
-
Then
Gĩkũyũ TTS
Build an open Gĩkũyũ text-to-speech model, evaluate naturalness with native speakers, and release a model and demo that can be used alongside the ASR work.
Aria na ithũĩ · Talk to us
This will take many voices. Add yours.
If you speak Gĩkũyũ, build machine learning systems, or want to bring this work to another language, we'd like to hear from you.