Hi,
I did not change the code, and training with cv-valid-train.csv and ted-train.csv as the training input dataset, use run-ted.sh as the parameter(just add cv-valid-train.csv). but the performance seems not good.
It is much worse than the performance you claimed.
Is there anything I can do to improve the performance?
Or Could you told me what is the training setting to get the performance of less 6% WER?
I found no data argument(speed change), no noise argument, no input feature normalization in the latest code. Are them helpful?
I remember the old version code has feature normalization.
after 10 epoch training,the test result is as below:
the training log is:
Thanks
Jinhong