Models · Speech to text
On RinggAI’s public Hindi benchmark it made the fewest word errors of five.
Word error rate
Why Fourier STT
It hears Hindi, English, Gujarati, Tamil and Telugu, and the English mixed into them.
01
Code-mixed Hindi and English, with names and numbers mid-sentence. Built for conversational, phone-line Hindi.
02
On Lahaja, 132 speakers from 83 districts it never trained on: 11.63% WER, against 16.55% for the open Qwen3-ASR model.
03
Built for conversational, phone-line Hindi and Hinglish, and served on our own engine and GPUs in India.
Benchmark
The same 10,000 utterances for every system, across six corpora.
| Detail | Fourier STT | Ringg Parrot | ElevenLabs | Sarvam | Deepgram |
|---|---|---|---|---|---|
| By dataset | |||||
| CommonVoice · 1,727 utterances | 12.78% | 12.62%lowest | 15.98% | 16.68% | 18.49% |
| FLEURS · 418 utterances | 9.89% | 10.36% | 9.74%lowest | 13.40% | 15.65% |
| IndicTTS · 100 utterances | 5.31% | 5.21%lowest | 10.57% | 10.32% | 8.65% |
| Kathbath · 1,929 utterances | 7.99%lowest | 9.46% | 12.14% | 13.32% | 14.18% |
| Kathbath noisy · 1,929 utterances | 9.47%lowest | 10.84% | 11.83% | 14.49% | 15.61% |
| MUCS · 3,897 utterances | 8.24%lowest | 8.41% | 8.97% | 10.50% | 16.03% |
| All 10,000 utterances | |||||
| Overall · 10,000 utterances | 9.12%lowest | 9.72% | 11.11% | 12.83% | 15.79% |
Pricing
Send a recorded call and get a transcript back, billed once the job completes.
Speech to text
₹18an hour of recorded audio³
Questions
Our own speech-to-text model, built for conversational, phone-line Hindi and Hinglish, and served on our own engine.
On all 10,000 utterances of RinggAI’s public Hindi benchmark, Fourier STT scored 9.12% word error rate, the lowest of the five systems scored. Ringg Parrot scored 9.72%.
Hindi, English, Gujarati, Tamil and Telugu, including English words mixed into them. The benchmark on this page is Hindi only. Bengali, Marathi and Kannada follow as each clears our own evaluation.
It is built for phone-line audio and runs inside our voice agents today. The benchmark on this page is public read and prompted speech, not phone calls, so we do not quote a phone-call accuracy figure from it.
Try it inside a voice agent in the sandbox today. For recorded calls, the transcription API (POST /v2/transcriptions) takes recordings of up to 2 hours and bills by the hour of audio. For live audio, book a call and tell us what your callers sound like.
Fifteen minutes with the team that builds it. Bring your script, or your hardest call.
Book 15 minutes
