I've also been focusing on ultra efficient local ASR models (for real time transcription) that can run comfortably on a web browser on low end phones/devices. it's actually crazy how capable it is already (https://starbright.quickthoughts.ca/) WER is like ~5 and it works on old phones that I've tested.
I've been interested in ASR models specifically for transcribing low-bandwidth low-quality radio recordings, are there any specifically focused on this angle in particular around?
rn just english but I have a welsh (cy) asr model cooking as well, though its still unusable (~25 WER). these tiny models are actually really cool example of distilling but also really really focusing on the engine in a way that takes advantage of crazy compression, decoding etc strategies
Agreed theres a lot between "editing raw weights" via LLM and a bunch of code/programming. But in my case, the technique itself was new to me, and worked for a small ASR model I've been training based on moonshine + claude being able to actually reason about what was causing loops. This is something I sure couldn't/wouldn't do manually.
reply