neha
0:000:06
नमस्ते, मैं नेहा बोल रही हूं Acme से। आपका order कल deliver होगा, बस address confirm कर लीजिए।
Rendered by Mira TTS · 2026-09-03
Mira TTS · Text to speech
In blind listening tests on Voice Arena, people prefer Mira TTS to Sarvam’s Bulbul V3. It speaks Hindi, English, Hinglish and Gujarati, and starts talking in under 300 ms even with 24 voices on one card.
Blind listener votes, Bradley–Terry Elo
+35 Elo ahead of Sarvam
24 September 2026 · 18 models rankedView on Voice Arena
Listen
Original model output, not edited. The text under each player is exactly what we sent.
0:000:06
नमस्ते, मैं नेहा बोल रही हूं Acme से। आपका order कल deliver होगा, बस address confirm कर लीजिए।
Rendered by Mira TTS · 2026-09-03
0:000:06
नमस्ते, मैं आशु बोल रहा हूं। आपकी EMI पांच तारीख को due है — क्या online pay करेंगे?
Rendered by Mira TTS · 2026-09-03
Hindi, English, Hinglish and Gujarati, switching from “नमस्ते” to “order confirm” in the same sentence.
First audio in 275 ms at p99 with 24 voices speaking at once, and three times the streams per GPU of the usual eight.
An OpenAI-compatible speech endpoint. Stream 48 kHz PCM as it renders, or ask for a WAV file.
Head to head
Voice Arena plays two voices saying the same line and asks listeners which sounds better, without telling them whose is whose. Thousands of those votes become an Elo rating.
| Category | Mira TTS | Sarvam Bulbul V3 | Mira’s lead |
|---|---|---|---|
| All categories | 1000 | 965 | +35 |
| General conversation | 1058 | 966 | +92 |
| Content creation & education | 1021 | 908 | +113 |
Overall: Mira TTS 1000 ± 26 from 440 votes; Bulbul V3 965 ± 9 from 3,882 votes. Bulbul V3 is narrowly ahead in customer support (991 against 978) and media & entertainment (973 against 963). Snapshot of 24 September 2026.
See the full leaderboardAn independent leaderboard for voice AI. Listeners hear two voices read the same text and pick the better one without knowing which model made which. The votes become a Bradley–Terry Elo rating for each model.
On Voice Arena’s Hindi leaderboard, Mira TTS scores 1000 Elo against 965 for Sarvam Bulbul V3 across all categories, and leads by 92 points in general conversation and 113 in content and education. Bulbul V3 is narrowly ahead in customer support and in media and entertainment.
First audio arrives in 275 ms at the 99th percentile with 24 voices synthesising at once on one GPU. You can stream PCM as it renders, so playback starts before the sentence is finished.
Hindi, English, Hinglish and Gujarati are live. The current voices are ashu, neha, shruti and sameer.
Get an API key, then call the OpenAI-compatible speech endpoint with model mira-tts. The quickstart has working examples.
Start on pay as you go and move to Business when you need more concurrency.