Research

Grounded in African linguistic data

We build on open datasets and pretrained multilingual models, then apply targeted fine-tuning to close the performance gap between high-resource and African language voice AI systems. All model weights, training configs, and evaluation results are published openly.

Projects

Mar 2026Published

asrwhisperswahilifine-tuning

SAUTI ASR v1: Fine-tuning Whisper for Swahili — from 27% to 13.5% WER

Off-the-shelf multilingual ASR gets Swahili wrong one word in four. We fine-tuned Whisper-medium on local Swahili data and halved that error rate to 13.5%. Here is what we learned.

Q2 2026Planned

translationstreamingswahiliearphones

Real-time earphone translation: bridging English and Kiswahili conversations

Two people, two languages, one conversation. We are building a real-time translation system where English and Kiswahili speakers can talk naturally through earphones — hearing each other in their own language.

Q2 2026Planned

voice-agentarchitectureivrtelephony

Designing a low-latency Swahili voice agent for telephony deployments

Architecture design for a full voice-turn IVR agent combining SAUTI TTS + ASR with a language model backend, optimised for sub-500ms response latency on African telephony networks.