Models · Speech to text
It hears the Hindi and Hinglish people speak on the phone, and gives the transcript back fast.
Why Mira STT
It hears Hindi, Gujarati, Marathi, Bengali, Tamil, Telugu, Kannada and Malayalam, and the English mixed into them.
01
Code-mixed Hindi and English, with names and numbers mid-sentence, each word kept in its own script.
02
On 158 Hindi phone clips sent from India, the final transcript came 82 ms after the caller stopped speaking (80th percentile).
03
Built for 8 kHz phone-line audio, and served on our own engine and GPUs in India.
The test
The same public clips through each company’s own API on 20 October 2026, scored the same way.
| Measure | Voice Infra | Sarvam |
|---|---|---|
| Errors on Hindi phone calls, word error rateLower is better | 18.3% 499 public phone recordings | 21.6% Saaras v4 |
| Final transcript after the caller stops, 80th percentileLower is better | 82 ms 158 Hindi phone clips, from India | 290 ms Realtime API |
| Errors on clean read Hindi, word error rateLower is better | 10.1% Kathbath read speech | 5.8% Saaras v4 |
Sarvam is better on clean read speech. The test, run on 20 October 2026 through each company's own API, is on our Sarvam comparison.
Pricing
Live audio by the minute, recorded calls by the hour, billed by the second.
Speech to text
₹0.30a minute of audio, over HTTP or the WebSocket³
₹15an hour of recorded audio, transcribed offline
Questions
Our own speech-to-text model, built for conversational, phone-line Hindi and Hinglish, and served on our own engine.
On 499 public Hindi phone recordings at 8 kHz, Mira STT made 18.3% word errors, against 21.6% for Sarvam’s Saaras v4 on the same clips. On clean read Hindi, Saaras does better: 5.8% against our 10.1%. Send twenty of your own calls and score them.
On 158 Hindi phone clips streamed from India, the final transcript came 82 ms after the caller stopped (80th percentile). Sarvam’s Realtime API took 290 ms on the same clips.
₹0.30 a minute of audio, live over the WebSocket or in HTTP clips of up to 120 seconds, metered by the second. Recorded calls of up to 2 hours cost ₹15 an hour, charged once the job completes; degraded and failed jobs are free. All prices exclude GST.
Hindi, Gujarati, Marathi, Bengali, Tamil, Telugu, Kannada and Malayalam, including English words mixed into them. The phone-call test on this page is Hindi.
Fifteen minutes with the team that builds it. Bring your script, or your hardest call.
Book 15 minutes