Orato ASR — Hindi v1
Full-parameter supervised fine-tune of Qwen/Qwen3-ASR-0.6B (audio tower frozen)
on ~1,000 hours of Hindi/English/Hinglish calling-domain audio (appointment
booking, support, insurance, sales).
Training
- Method: full SFT, decoder + projector updated, audio encoder frozen
- Params: ~0.8B total, ~0.6B trained (audio tower frozen)
- Data: ~1,000 hours, calling-domain Hindi/Hinglish
- LR 1e-5, cosine schedule, warmup 0.03, max_grad_norm 1.0, 1 epoch, 1x H100
Evaluation (base vs. fine-tuned, same held-out sets)
| Benchmark | Base WER | Fine-tuned WER | Relative improvement |
|---|---|---|---|
| Kathbath | 15.24% | 11.49% | 24.6% |
| Gramvaani | 39.07% | 37.66% | 3.6% |
| Lahaja | 25.09% | 18.68% | 25.5% |
| Fleurs | 19.12% | 16.58% | 13.3% |
Fine-tuning improved WER on every benchmark tested, most on the benchmarks closest to the real product domain (telephony, accents, conversational speech).
Comparison with the Hindi ASR market
| System | FLEURS | CommonVoice | IndicTTS | Kathbath | Kathbath-Hard | Gramvaani | Lahaja |
|---|---|---|---|---|---|---|---|
| Orato ASR | 16.98 | 20.63 | 13.23 | 11.49 | 13.21 | 37.66 | 18.68 |
| Azure STT | 24.30 | 14.60 | — | 13.60 | 15.10 | 42.30 | 28.60 |
| Deepgram | 17.14 | 20.04 | 13.65 | — | — | — | — |
| ElevenLabs Scribe | 11.99 | 16.59 | 16.06 | — | — | — | — |
| Google Chirp | — | — | — | — | — | — | 22.30 |
| Google STT | 19.40 | 20.80 | 18.30 | 14.30 | 16.70 | 59.90 | — |
| IndicASR M1 (Conformer-L) | — | — | — | — | — | — | 19.40 |
| IndicWav2Vec | — | — | — | — | 16.20 | 42.10 | — |
| IndicWhisper | 11.40 | 15.00 | 7.60 | 10.30 | 12.00 | 26.80 | — |
| MMS (300M) | — | — | — | — | — | — | 34.40 |
| Nvidia Conformer-M | — | — | — | 14.00 | 15.60 | 41.30 | — |
| Sarvam Saarika | 16.00 | 18.21 | 15.37 | — | — | — | — |
| Whisper large-v3 (1.55B) | — | — | — | — | — | — | 32.40 |
Sources: AI4Bharat Vistaar benchmark suite (IndicWhisper, Nvidia Conformer-M, IndicWav2Vec, Google STT, Azure STT on FLEURS/CommonVoice/IndicTTS/Kathbath/ Kathbath-Hard/Gramvaani); AI4Bharat Lahaja benchmark; third-party published WER for Sarvam Saarika/Deepgram/ElevenLabs Scribe on FLEURS/CommonVoice/IndicTTS. Normalization may differ slightly across sources — treat as directional.
Loading
import qwen_asr, torch
wrapper = qwen_asr.Qwen3ASRModel.from_pretrained(
"tryorato/orato-asr-hindi-v1", dtype=torch.bfloat16, device_map=None,
attn_implementation="sdpa",
)
wrapper.model = wrapper.model.to("cuda")
result = wrapper.transcribe(audio=(wav, 16000), language="Hindi")
Must load via qwen_asr.Qwen3ASRModel, not transformers.AutoModel — this
is the non--hf wrapper checkpoint; loading it as a plain Transformers model
will silently misbehave.
Intended use / limitations
Beta. Evaluated on Hindi calling-domain audio; not evaluated outside that domain. Access is currently private.
Data rights
Training data includes a mix of public research datasets — Rasa Hindi, Gram Vaani, MUCS — used under their respective published licenses, plus proprietary Hindi/Hinglish calling-domain audio collected by Orato.
- No raw audio, transcripts, or personally identifiable data are included in this repository — only model weights are published.
- Commercial-use licensing confirmation for the public source datasets, and consent/PII review for the proprietary calling data, are in progress. Released for internal/beta evaluation only until that review completes.
- This repository does not convey redistribution rights to any underlying training data.