Feature Request/Advice : Fine Tuning for Medical Use Case?

[GaleiqTesting]

Hey there, thanks for making this implementation for MLX. Really appreciate it. Your projects help open source significantly!
Comparing it to Whisper Turbo V3 (local as well).

Parakeet:
Pros;

• Much Less resource intensive
• Significantly Faster - especially on longer audio files

Cons

• Less accurate on medical recordings as compared to Whisper large-v3-turbo
• Non significant time difference for audio around the range 0-5min.

Feature Request: Simple Way to drop in a dataset and fine tune parakeet?

Thanks a lot!
(Tested Combinations: both wav and m4a, with and without chunking, fp32 and bf16)
(Context; I'm a medical doctor and not very well versed in coding - mainly just use Cursor/RooCode + Gemini/Grok/Claude to help experiment and trial stuff.)
(Specs: Macbook 14 Pro M1 Max 64GB RAM)

[kth8]

Did you use the default 0.6B model? Try the larger 1.1B ones.

[kth8]

I have been comparing this medical recording with whisper-large-v3-turbo and parakeet-tdt-1.1b, both seem to have done well, but I'm no medical doctor.
https://www.voxtab.com/downloads/audio-files/VOXTAB_Medical_audio.mp3

[senstella]

Hello, thank you for trying out! While fine-tuning capacity in MLX isn't quite planned (the original NeMo implementation already excels here), I'll see what I can do at inference time to improve accuracy using a dictionary of words!

[deadcoder0904]

Would love this too. I am in web development & sometimes transcribe videos of AI Coding or SEO so some terms are miswritten.

I would like to drop in a dictionary too to fix it. I think https://aquavoice.com/ is the best in class that solves hard problems like this in voice. moree at https://aquavoice.com/avalon-api