Speech AI for Indian languages in health voice assistants
Sarvam, AI4Bharat, Bhashini and global providers compared for Hindi, Marathi, Tamil and other Indian languages, and how to test speech models on your own callers.
- India
Key takeaways
- Sarvam reports that its Saaras V3 model beat Gemini, GPT-4o Transcribe and Deepgram Nova-3 on the IndicVoices and Svarah benchmarks.
- Sarvam says Saaras V3 was trained on more than 1 million hours of Indian speech.
- AI4Bharat, at IIT Madras, publishes open Indian-language speech models and benchmarks such as IndicVoices.
- Vendor benchmarks are not your callers. Measure word error rate on 100 to 300 real, consented recordings per language.
- For health, latency matters as much as accuracy: replies slower than about a second break phone conversations.
What are the main speech AI options for Indian languages?
| Provider | Type | Strengths | Consider |
|---|---|---|---|
| Sarvam (Saaras, Bulbul) | Commercial APIs, India-built | Reports leading Indian speech accuracy; Indian voices for text-to-speech | Accuracy figures are vendor-reported |
| AI4Bharat | Open models and datasets from IIT Madras | Self-hostable, wide language coverage, public benchmarks | You run and scale it |
| Bhashini | Government language platform | Public translation and speech services | Check service levels for production |
| Global clouds (Google, Microsoft, AWS) | Commercial APIs | Mature tooling, many languages | Indian accents and code-mixing vary |
| Specialist voice platforms | Commercial | Low-latency real-time pipelines | Indian language depth varies |
Why do benchmark results differ so much?
Benchmarks use different audio: read speech, conversational speech, studio or phone quality, single language or code-mixed. A model can win on clean read Hindi and lose on noisy Hinglish phone calls. Recent results also mostly come from the vendors who built the models.
How should a health team test speech models?
- Record 100 to 300 consented calls per language from real users, with identifiers removed.
- Transcribe them carefully by hand, including medical terms.
- Measure word error rate overall and on medical words, symptoms and drug names.
- Measure end-to-end latency on a real phone line.
- Re-test every time a provider ships a new model.
What else matters for health voice assistants?
Turn-taking that tolerates pauses, gentle confirmation of critical details such as dates and doses, a fast path to a human for crises, and text-to-speech voices that sound local and trustworthy.
Frequently asked questions
Which speech-to-text is best for Hindi health calls?
There is no universal answer. Indian-focused models currently report the lowest error rates, but results vary by accent, background noise, phone line quality and medical vocabulary. Run the same recordings through two or three candidates and compare.
Can voice AI run without sending audio to a foreign cloud?
Yes. Open models from AI4Bharat and others can be self-hosted, and Indian providers offer in-country processing. Check where audio and transcripts are stored and for how long.
Sources
- Saaras v3 beats Gemini, GPT-4o on Indian speech benchmarks, says Sarvam AIBusiness Standard · 2026-02
- Introducing Saaras v4Sarvam AI · 2026
- AI4BharatIIT Madras
- BhashiniMinistry of Electronics and IT, India
Found something out of date? Write to contact@brandlabs.app and we will fix it.