Int8 issues, I fixed it I think
#3
by Olicorne - opened
Hi,
I think I found a better way to make an int8 version of parakeet that greatly lowers the WER at the cost of a small size bump.
I documented the whole thing here: https://huggingface.co/Olicorne/parakeet-tdt-0.6b-v3-smoothquant-onnx
Feedbacks welcome!