Semi-tight covariance matrices implementation in MASPER HMM training procedure

摘要

The article discuses aspects of incorporating shared linear transformations to implement semi-tight covariance matrices into MASPER HMM training procedure. The concern is on heteroscendic linear discriminative analysis (HLDA) applied to speech features. Next main implementation issues and necessary modifications to the standard MASPER training procedure are introduced. Finally an evaluation of the suggested and implemented changes is accomplished. Apart of that large vocabulary word loop test (LVWLT) has been designed and implemented in order to provide finer system evaluation. All experiments have been executed on Slovak part of MobilDat training database. Achieved results show that incorporation of semi-tight matrices is beneficial in the terms of word error rates (WER) for wide range of test scenarios and models. However, it is for the sake of the increased number of free parameters, sensitivity of complex models to the overtraining and longer training times.