Technology
Fine-tuned automatic speech recognition for African languages — starting with Kiswahili customer experience (CX) — with measurable gains over out-of-the-box baselines.
We fine-tune from OpenAI Whisper-small — a strong multilingual baseline — then adapt with telephony-style augmentation and domain data as partners share it.
How it works
Data collection
Today we train from strong public and licensed Kiswahili corpora. In parallel we collect and partner for Kenya CX telephony and code-switched speech — accents, product vocabulary, and call types that match production — so each pilot improves on audio like yours.
Fine-tuning
Starting from Whisper, we adapt checkpoints to your languages and acoustic conditions. Evaluation uses normalized references and word error rate so improvements are measurable, not anecdotal.
Deployment
Tuned models sit behind a production API you can call from existing workflows. You control rollout: compare baseline versus tuned in the same request when you need side-by-side assurance.
Continuous improvement
New partner clips, validator feedback, and production traffic feed the next training cycle — closing the loop from capture to better transcripts over time.
Accuracy
On the public FLEURS Kiswahili (Kenya) read-speech evaluation set, our production checkpoint reduces word error rate from 106% (out-of-the-box Whisper-small) to 19.5% — about an 82% relative improvement. These figures are benchmarks on that set, not a guarantee for every recording condition. Real call-centre word error rate is measured on your held-out audio during a pilot.
Compare mode
Every API call can return baseline and tuned transcripts in one response, so you can see the improvement on your own audio before a rollout.
Supported languages
Kiswahili is the primary focus today, with emphasis on Kenyan customer-experience speech — including fintech and mobile-money vocabulary as we expand partner data. Our roadmap includes additional East African low-resource languages as data partnerships and evaluation coverage grow.