June 2026
Kikuyu TranslateGemma 4B: Faster, Smaller, Better
How we replaced our 12B translation model with a faster TranslateGemma-4B rsLoRA model — improving BLEU to 21.93 and making C-elo Translate feel usable on the web.
researchtranslation
Real-time voice and text AI for African languages. We build streaming ASR, translation, voice chatbots, and text-to-speech models so businesses can serve customers in Kikuyu, Dholuo, Kalenjin, Kamba, and more.
Our 12B TranslateGemma model achieved a 758% improvement over zero-shot; our newer 4B model improves quality to 21.93 BLEU while serving faster. Our real-time ASR models stream Kikuyu and Dholuo transcripts as users speak, and our voice chatbot lets you converse with AI in spoken Kikuyu.
About Us
C-elo Labs builds production AI for underserved African languages. We fine-tune open-source models on curated African Languages datasets and deploy them as real products.
Our work spans translation (TranslateGemma), real-time ASR for Kikuyu and Dholuo, voice chat (ASR + LLM + TTS), and Speech-to-Speech AI (Mimi codec), with expansion to Kamba, Kalenjin and other Low Resource languages on the roadmap.
Learn more about usSee It In Action
Watch C-elo models transcribe, translate, speak, and converse in African languages — live demos from our platform.
What We Build
Production AI tools for African languages — from real-time transcription and translation to spoken voice interaction.
Translation
LiveEnglish → Kikuyu translation: our 12B model achieved 19.61 BLEU — a 758% improvement over zero-shot — and our current 4B V7 reaches 21.93 BLEU with faster responses.
Real-time ASR
Live BetaStreaming speech-to-text for Kenyan languages. Kikuyu and Dholuo ASR are live in C-elo AI, with cache-aware Nemotron-style streaming built for low-latency transcription.
Voice Chatbot
BetaSpeak in Kikuyu and hear AI respond in Kikuyu. Cascaded pipeline: MMS ASR → LLM → MMS TTS, running on serverless GPUs.
Text-to-Speech
LiveConvert Kikuyu text to natural spoken audio using Meta's MMS-TTS model, adapted for Kikuyu phonetics.
Navigation (C-elo Maps)
LiveTurn-by-turn driving directions with voice guidance in Kikuyu — the first navigation app in an African indigenous language.
Speech-to-Speech AI
In DevelopmentEnd-to-end voice model using the Mimi neural codec — aiming for real-time, full-duplex Kikuyu conversations. Stage 1 codec adaptation complete.
Development Progress
Translation88%
DatasetData
Fine-tuneTrain
DeployDeploy
Human EvalEval
Voice AI63%
ASR + TTSASR/TTS
Mimi CodecCodec
StreamingStream
End-to-EndE2E
Streaming ASR75%
Kikuyu LiveKikuyu
Dholuo LiveDholuo
Kalenjin EvalKalenjin
API AccessAPI
Text Chatbot13%
Data PipelineData
LLM Fine-tuneTrain
DeployDeploy
Multi-langMulti
Complete
In Progress
Upcoming
C-elo Maps
LiveThe first navigation app with voice guidance in an African indigenous language. Search, navigate, and hear directions — all in Kikuyu.
Turn-by-Turn Navigation
Full driving directions with step-by-step maneuvers, route alternatives, and automatic rerouting.
Kikuyu Voice Guidance
Hear every navigation instruction spoken in natural Gĩkũyũ — powered by C-elo's voice AI.
Built for Africa
Optimized for Kenyan roads and locations. More African languages launching soon.
Available soon on Android
Starting with Kikuyu; Luo, Kalenjin, and Kamba coming next.
iOS in development.
Download APPv1.0.0
Research
Our research focuses on fine-tuning LLMs for African Low Resource languages using advanced techniques like LoRA, QLoRA and synthetic data generation, plus adapting streaming ASR and Speech-to-Speech models for African speech.
We collaborate with initiatives like African Next Voices and Google's WAXAL to create robust datasets and push the boundaries of low-resource language AI.
Our first streaming ASR paper is now available on arXiv.
Latest Updates
February 2026
Adapting the Mimi Neural Codec for Kikuyu Speech
How we fine-tuned Kyutai's Mimi codec on 750+ hours of Kikuyu audio with a custom pitch-preservation loss — the foundation for real-time voice AI.
researchvoice-ai
January 2026
Building Kikuyu TranslateGemma: From 2.29 to 19.61 BLEU
How we fine-tuned Google's TranslateGemma-12B for English→Kikuyu translation — achieving a 758% improvement through systematic LoRA tuning and production deployment.
researchtranslation