Phone error rates below ~50%

[dspavankumar]

Phone error rates below ~50%

[bjtommychen]

Hi, Kumar
Thanks for your clue !
I think the problem is resolved now.
Although the Keras evaluate() report acc ~50%, the RESULTS script print out WER ~25%.
That means I got the almost same WER as yours.
Thanks again for your project. It help me lot when dig in Kaldi.

[dspavankumar]

Hello, can you verify the points below?

1. What is the WER you get for tri2? If it is worse than 50%, then something went wrong in the GMM training or feature extraction.
2. Make sure tri2 directory (the one you provide as 'gmm' to train.py) was indeed used to generate the alignments you are using to train the DNNs. Have you changed the tree file or retrained in the same directory after generating alignments?

I don't have the log file anymore, I used a different machine to generate the results.