Lessons from deploying voice AI in Tanzanian hospitals

Code-switching is the norm, not the exception

In the cardiology ward at Muhimbili, a single doctor-patient consultation routinely shifts between Swahili, English medical terminology, and one of several regional languages. Off-the-shelf ASR systems collapse under this load.

Three things we changed

1. Mixed-vocabulary acoustic modeling rather than language-detection switching
2. Domain-specific fine-tuning on 1,200 hours of clinical audio (consented and de-identified)
3. On-device inference so patient audio never leaves the hospital network

Results so far

Word error rate dropped from 41% (baseline) to 12% on real ward recordings. More importantly, doctors stopped switching the recorder off mid-consultation - the trust signal we actually care about.