Lantian Li 2015-04-08

Weekly Summary

1. Addtional experiments are made for semi text-independent SV based on d-vector.

Results show that phone posteriors (without utterance CMVN but make Global CMVN)

combined with Fbanks (without utterance CMVN but make Global CMVN) as DNN input feature

performs the best.

2. Addtional experiments are made for text-dependent SV on d-vector. Aiming to compare the performance between

CNN and DNN.

Results show that the performance of d-vectors extracted from the traditional DNN are better than CNN's on this text-dependent SV task.

3. Prepare for the cohort-based classfication task in score-level on GMM-UBM system.

Next Week

1. Go on the task 2, 3.