How to inference with s4 decoder?

[Haoxiang-Hou]

When I use s4 decoder to train in Librispeech, asr1, the loss is very well.
However, when I inference with s4 decoder, the WER is very bad. And the inference beamsearch CER is much bigger than training CER and CER-CTC. It is strange.

When I use s4 decoder to train on Librispeech_clean_100, asr1, the valid CER is 0.076, and ctc CER is 0.086.

When I use s4 decoder to inference, the beamsearch CER of dev_clean is 16.7%, even worse than training.

[m-koichi]

Hi, thanks for your report.
I confirmed that I was able to run the S4 decoder training and inference with the latest commit successfully.
Could you share your training and inference configurations?
Also, could you check the Transformer decoder inference as well?