Zhiyuan Tang 2016-04-18

Last week:

1. enhancing the joint model with SWBD focusing on speech recognition, shows improvement[1];

2. problem remains that when the WSJ was reduced to be 8k, the advantage of joint training disappeared at least for speech recognition. (comment later: WSJ was reduced to 8k by mistake, so the pipeline needs to be reconducted, conlusion 1 still stands).

This week:

1. find the reason why joint training failed on 8k WSJ;

2. more experiemnts for refining the joint model, such as enhancing the enhanced model again with speaker data;

3. following ICASSP 16.