LLM post processing

[AlexandderGorodetski]

Hello guys,

I am working with my inhouse data, which contains 5000 hours of noisy audio.

I found that applying of Chat GPT as error correction post processing in some benchmarks can decrease WER up to 10% absolute.

I am pretty sure that for ASR error correction we do not need so big model as Chat GPT 4.o.

My question is if you do not plan to add recipe for LLM training that can be used for error correction.

Thanks a lot, AlexG.