Voice AI for low-literacy health access.
Latency, vernacular STT/TTS, and when voice beats chat — lessons from rebuilding Myna Voice AI for rural women’s health support.
·8 min read
TL;DR
Voice is the right interface when typing fails — but only if latency is near-instant and vernacular STT/TTS is solid. Brandlabs rebuilt Myna Voice AI on Pipecat with Sarvam for vernacular speech and cut multi-second delays that made phone health support unusable.
When should a health product choose voice over chat?
When users already send voice notes, literacy or script switching is a barrier, or hands-free phone calls are the only reliable channel. Voice is not a novelty feature — it is accessibility. If your users speak more than they type, chat-first is an exclusion design.
Myna Voice AI extended the Myna Bolo ecosystem for women who prefer speaking — especially low-literacy rural users. The product insight was obvious; the engineering insight was latency.
A 3–4 second gap after each turn destroys conversational trust. People hang up. They assume the system is broken. They do not “wait for the LLM.”
What technical choices actually fix voice health assistants?
Streaming pipelines, vernacular STT/TTS matched to the dialects you serve, intent routing before long generations, and RAG that stays within approved medical knowledge. Brandlabs moved Myna Voice onto Pipecat with Sarvam AI for speech and OpenAI for orchestration — prioritizing turn speed as a product requirement.
Pipeline shape that worked:
incoming audio → vernacular STT → jargon filtering → LLM + RAG → vernacular TTS → phone delivery
Optimize the critical path ruthlessly. Fancy agent graphs that add seconds are regressions, not architecture.
How do evals change for voice?
Score transcript quality, end-to-end latency, barge-in behavior, and safety on spoken phrasings — not only text prompts. A text suite that passes while voice emergency phrases fail is a false green.
Add to your suite:
- Locale audio fixtures (or transcript variants from real anonymized calls)
- Latency budgets per turn
- Safety cases in spoken form
- Analytics on completion rate, duration, and repeat engagement
We instrumented voice interactions for confidence, fallback, and language splits — because you cannot improve what you only demo.
Frequently asked questions
Is voice always better for rural health?
No. Noisy environments, shared spaces, and stigma can make voice worse. Offer chat and voice; let the context choose.
Which matters more — model quality or latency?
For phone turns, latency is table stakes. A slightly simpler answer that arrives now beats a perfect answer that arrives after the user leaves.