140707 -Xuewei Zhang

From cslt Wiki
Jump to: navigation, search

Last week 1.in order to find the reason that the DAE does not work well,realign recording(done) 1.train mpe_adaptation_dae_cmvn-g_splice(done) 2.study DAE based on MATLAB,but I have a problem how to input batches of matrix to train a nnet.When I input one matrix,matlab warned : matrix is too large.When "for cyclic sentence" runs the first batch of matrix,the net is trained once,however,when "for cyclic sentence" runs the second batch of matrix,the net is trained again.

This week 1.solve the problem in the last week,but the mothod of finding transformation does not work well so that we should search another method. 1.realize dereverberation based on Lasso and filter(SINGNAL CHANNEL REVERBERANTION SUPPRESSION BASED ON SPARSE LINEAR PREDICTION) 2.read papers to find another effective approach to realize dereverberation