“Zhiyuan Tang 2016-05-30”版本间的差异

2016年6月6日 (一) 00:39的最后版本

Last week:

1. all partial models will be generated in one day, more results in[1],

conclusion: Enhancing on the part of spk seems not help speech while enhancing speech helps spk. It's consistent with that for partial when data for spk increases and data for speech keeps the same, speech performs worse while spk becomes better.

Test on wsj_test and eval2000, vad is important for baseline sre, while for joint, no vad or lower threshold for vad are better.

2. bilingual, results in[2];

This week:

1. more improvement on bilingual, such as p+r as interaction info;

2. paper initialization for joint training of journal version.