Have anyone tried 8-bit quantization for Conformer?

[grazder]

I have a large Conformer and I'm exporting it with ONNX in fp16. I'm interested in its quantization.
Have you tried this? Maybe with TensorRT or other frameworks? Can you recommend any working recipes for it?
If it worked for you, what kind of speedup did you get? What is WER reduction?

I saw https://github.com/kssteven418/Q-ASR for ASR solutions.

So, I'm interested in your experience with quantizing the Conformer/other ASR models. Can you share it please?