Request support for omi-med-stt-v1 (Fine-tuned Parakeet 0.6B) for Medical & Clinical Transcription

[koloved]

https://huggingface.co/omi-health/omi-med-stt-v1

Since Handy already supports the CPU-optimized Parakeet V3 model, I would like to request adding support for Omi Med STT v1. It is a brand-new, open-weight medical ASR model fine-tuned from NVIDIA's Parakeet TDT 0.6B v2.

According to recent benchmarks on clinical speech (including dictation, medication reviews, and medical dialogues), the base Parakeet model struggles heavily with specialized terminology. Omi Med STT v1 completely outperforms it in medical contexts while maintaining the same 0.6B model speed:

1. Massive Drop in Medical Error Rate (M-WER): Omi Med STT slashes the medical word error rate from 8.01% (Parakeet v3) down to 2.37% (Omi NeMo) / 3.20% (Omi GGUF).
2. General Accuracy (WER): Almost halves the overall error rate, dropping it from 15.26% down to 8.30%.
3. Drug Name Precision: It dramatically reduces hallucinations and incorrect drug spelling. Drug M-WER drops from 9.50% (Parakeet v3) to 4.75% (Omi).

I am not a medical professional or a doctor, but it seems like this would be great to add for people working in this field. But the only thing that will need to be converted is it into int8 format.
Can anyone who worked on the conversion of this model clarify how it was done? I could do this.

[cjpais]

Will be included as part of transcribe.cpp, thanks for pointing it out