validation wer of training mode and test mode are different.

[burin-n]

I trained the model and got a good validation WER.
However, I got very poor decoded results when I tried to load that checkpoint and ran it in --test mode.

Do you have any suggestion on this?

[kouohhashi]

I have a similar experience. have you find a solution?

[burin-n]

There is a problem with the saved checkpoint. I'm not sure what happened and how was it happen.
As I found, the weights of latest.pth were not actually come from the last epoch.
I retrained the model and never found this problem again.

[WAIML]

Hi All,
Thanks for author's work!
I trained the models recently with librispeech train-cleain-100, and the WER is very low during training:
INFO] Saved checkpoint (step = 80.0K, wer = 0.30) and status @ ckpt/Librispeech_subwords_1000/best_att.pth
when tested the performance with the best_att.pth, on test-clean or dev-clean, the performance is very poor, WERs were about 80%.

@kouohhashi Have the problem of high-reference-WER (about 80%) being resolved?
@burin-n After you retrained the model did you get good performance, i.e., low WER?

Any updates for the WER for this repo, or any solutions for the problem?
If this one does not work, do you have any recommendations for a simple ASR package works?

Thank you very much for your helpful information!

Regards,
Willy