Zhiyuan Tang 2015-03-23

From cslt Wiki
Jump to: navigation, search

Last week:

got a better understanding of ASR with some papers;

did some easy analysis about the weights of DNN already learned from Kaldi to see if some intersting things could be found and more regularization could be added to train a better DNN;

the CONTEXT into DNN may be not enough, as the central frame is not highlighted and its neighbors should not share the same importance. Analysis of the first-layer weights may give some insight to improve the CONTEXT.

This week:

do more analysis about the learned weights;

improve the CONTEXT if possible, at least give it some small changes and contrast its performance with previous results.