Voice quality is the single biggest reason an AI calling pilot dies. A human picks up, hears something off — robotic, wrong accent, weird cadence — and hangs up before the disclosure even finishes. We ran a structured blind test of the three TTS providers we use in production. Here's what we learned.
Test setup
- Three providers: Sarvam AI, Smallest AI, ElevenLabs.
- Six languages: Hindi, Tamil, Telugu, Marathi, Bengali, English (Indian).
- Eight evaluators across India — native speakers of each language.
- 12 sample utterances per language: greeting, intent capture, objection handling, closing.
- Each utterance synthesized by all three providers, presented blind in randomised order.
- Evaluators rated naturalness (1-5), accent fidelity (1-5), and would-they-stay-on-the-call (yes/no).
Naturalness scores (5-point scale, higher is better)
Hindi
- Sarvam: 4.4 — clearly the strongest. Indian accent baseline, no English-accented Hindi.
- Smallest: 3.8 — clean output, occasional slight English-accented inflection.
- ElevenLabs: 3.2 — usable but evaluators flagged 'this is an Indian voice trained on English data'.
Tamil
- Sarvam: 4.6 — best in class. Native cadence, dialect-aware.
- Smallest: 3.4 — passable, some pronunciation drift on retroflexes.
- ElevenLabs: 2.8 — evaluators said 'sounds like Hindi-speaker reading Tamil'.
English (Indian)
- Sarvam: 4.0 — solid Indian English, slight monotone on long utterances.
- Smallest: 4.2 — best-in-class for Indian English specifically.
- ElevenLabs: 4.5 — best overall for English; the 'Olivia' and 'Ethan' voices we use on the Premium tier are ElevenLabs-powered.
Latency (TTS-only, p50, ms to first audio)
- Sarvam: 280ms
- Smallest: 220ms (best)
- ElevenLabs: 380ms
Anything under 400ms is conversational; the human ear doesn't notice. All three pass the bar. Where it matters: full agent-loop latency (STT + LLM + TTS) — we'll cover that in a follow-up post.
Stay-on-the-call rate
The most consequential number. Across 96 evaluations:
- Sarvam (Indic only): 91% would stay on the call.
- Smallest (Indic only): 76% would stay.
- ElevenLabs (Indic only): 58% would stay.
- ElevenLabs (English only): 88% would stay.
What surprised us
Three things:
- The Bengali results were closer than we expected — Sarvam still won, but Smallest's Bengali was within 0.3 points. Bengali might be a place where alternatives become viable.
- Tamil was the language where the gap was widest — and where Cualify customers are most demanding. If you're targeting Tamil-speaking SMBs, you cannot afford a non-Sarvam TTS.
- Punjabi (which we tested on the side, n=4) was rough across all three. We're holding off on shipping Punjabi voices until either Sarvam or a new entrant raises the bar.
Methodology caveats
Three caveats worth flagging:
- Evaluators were Cualify-network volunteers, not paid. Selection bias toward urban/tech-adjacent listeners.
- We didn't test against a baseline of human voice agents in the same pipeline. That's a follow-up.
- The TTS landscape moves fast — these scores are from March 2026. Smallest in particular is improving rapidly. Re-test in 6 months.
Want the raw evaluator scores? Email [email protected] and I'll send the spreadsheet. We're publishing this stuff because the Indian AI voice ecosystem is too small for vendors to fight blind.
Read more