🎙️ Srota: Hinglish Speech Recognition
Srota transcribes Hinglish the way people actually speak it, keeping English in Latin and Hindi in Devanagari (मेरा favourite festival Diwali है), instead of mangling code-switched speech into all-Devanagari transliteration. Record or upload a clip to try it.
Srota is a full-parameter fine-tune of Qwen/Qwen3-ASR-0.6B, trained on the union of HiACC (conversational) and OpenSLR-104 (technical tutorials). Full model card: moorlee/qwen3-asr-0.6b-hinglish.
Built by the team behind susrota.com, a voice-dictation tool (currently English; Srota will power its Hinglish support).
Word Error Rate (lower is better):
| Model | HiACC test (conversational) | OpenSLR-104 test (tutorials) |
|---|---|---|
| Base Qwen3-ASR-0.6B (zero-shot) | 24.73% | 50.66% |
| Srota | 15.85% | 35.06% |
Running on a free CPU Space: the first transcription after the Space wakes may take ~10-20 s.