Jacobian adaption of acoustic models for noise robust speech recognition

Abstract

In this thesis, a differential adaptation method, namely Jacobian Adaptation (JA), is researched for compensation of additive noise during recognition. Three versions of Jacobian adaptation were implemented. JA was compared to Parallel Model Combination (PMC) and Cepstral Mean Normalization (CMN). According to the results obtained in this thesis, JA showed comparable performance to CMN, but PMC outperformed JA. It has to be noted, however, that the implementation of PMC has higher computational cost than JA. /Kir10