Neural Machine Translation
We fine-tune translation models like Google's TranslateGemma for English↔Kikuyu using Parameter-Efficient Fine-Tuning (LoRA). Our first 12B model achieved 19.61 BLEU — a 758% improvement over zero-shot performance on 30,430 curated sentence pairs. Our current production 4B model improves this to 21.93 BLEU and 42.87 chrF++ while serving faster on the web.
Key areas: LoRA and rsLoRA optimization, regularization tuning, BLEU/chrF++ evaluation, production inference and latency optimization
Streaming Automatic Speech Recognition
We adapt cache-aware streaming ASR models for Kenyan languages so users can see transcripts update while they speak. Our current C-elo AI ASR demos include Kikuyu and Dholuo, with Kalenjin evaluation and refinement in progress. The research stack focuses on low-latency audio chunking, language-specific prompt conditioning, orthography cleanup, and stable partial transcript updates for live microphone use.
Key areas: Nemotron-style FastConformer-RNNT adaptation, cache-aware streaming, WER/CER evaluation, streaming inference
Speech-to-Speech Models
We are building end-to-end voice AI using the Mimi neural codec adapted for Kikuyu tonal fidelity. Our Stage 1 codec adaptation is complete (79.3M params, 1.1 Hz pitch error), with streaming inference and full-duplex conversation in development.
Key areas: Mimi codec adaptation, pitch-preservation loss, cascaded ASR→LLM→TTS pipeline
Dataset Engineering
We use the African Next Voices corpus (750+ hours of Kikuyu audio) and Google's WAXAL TTS dataset (~9 hours studio quality) alongside language-specific ASR evaluation sets to train, test, and compare robust speech and translation models.
Key areas: Audio preprocessing, noise augmentation, transcript normalization, multi-source dataset curation
Inference & Deployment
We deploy production models behind authenticated APIs and WebSocket streaming endpoints, with reliability and latency controls designed for interactive applications.
Key areas: Streaming inference, authenticated APIs, reliability, observability
