Medical speech recognition for Indian English, Hindi, Marathi, Kannada and Telugu.
An ASR model for doctor-patient conversations, including code-mixed speech with English drug names inside Indic sentences. A Whisper large-v3 encoder feeds a MedGemma 4B decoder, tuned further with GRPO. Lowest semantic WER in every language against nine production ASR systems.
Benchmarks
| Metric | Value | Dataset |
|---|---|---|
| Semantic WER, Indian English (lower is better) | 9.15 | Eka medical conversations |
| Semantic WER, Hindi (lower is better) | 18.16 | Eka medical conversations |
| Semantic WER, Marathi (lower is better) | 31.61 | Eka medical conversations |
| Semantic WER, Kannada (lower is better) | 29.24 | Eka medical conversations |
| Semantic WER, Telugu (lower is better) | 24.77 | Eka medical conversations |
Install
Download from Hugging Face
hf download ekacare/parrotlet-a-2.5-pro Links
Topics
asrspeech-recognitionmedical-transcriptionmultilingualcode-mixingwhispermedgemma