Trelis Tara
Frontier Hindi transcription with open weights.
State-of-the-art aggregate accuracy on the AI4Bharat Vistaar Hindi benchmark suite, with a native mixed-code mode for Hindi-English (Hinglish) speech.
Built for real Hindi speech
Tara handles the audio Hindi ASR actually meets in production: telephony, noisy environments, spontaneous conversation, accented speakers, and Hindi-English code-switching.
State-of-the-art Hindi
Best mean accuracy across the 7 Vistaar Hindi benchmarks, ahead of leading commercial Hindi ASR systems.
Native mixed-code mode
Writes English in Latin script and Hindi in Devanagari, the way real Hinglish is written. On pure Hindi audio it neither degrades accuracy nor forces transliteration.
Open and compatible
Apache-2.0 open weights, drop-in compatible with whisper-large-v3 tooling in 🤗 Transformers, plus a managed commercial API.
Vistaar Hindi benchmark (WER, lower is better)
| Benchmark | Trelis Tara | Sarvam Saaras-v3 | ElevenLabs Scribe-v2 |
|---|---|---|---|
| Kathbath (clean read) | 9.34 | 9.72 | 9.60 |
| Kathbath-hard (noisy) | 10.82 | 10.55 | 11.11 |
| MUCS | 10.79 | 9.69 | 10.93 |
| GramVaani (telephony) | 21.03 | 23.00 | 26.93 |
| IndicTTS | 9.46 | 10.38 | 13.17 |
| CommonVoice-hi | 12.51 | 12.88 | 13.44 |
| FLEURS-hi | 10.46 | 10.03 | 11.33 |
| Mean (7 Vistaar sets) | 12.06 | 12.32 | 13.79 |
Corpus WER under a Hindi-aware text normalizer, clips ≤ 30 s, identical references for all systems. Commercial-system results are measured by Trelis under the same protocol, not vendor-reported. Full details on the model card.
Commercial API or open weights
Start immediately through Trelis Router, or download the model from Hugging Face for self-hosted use.