ITERATIVE CONSTRAINED MLLR APPROACH FOR SPEAKER ADAPTATION

摘要

In this paper an effective technique for speaker adaptationrnon the feature domain is presented. This technique startsrnfrom the well known maximum-likelihood linear regressionrn(MLLR) auxiliary function to obtain the constrainedrnMLLR transformation in an iterative fashion. The proposedrnapproach is particularly suitable to be implemented onrnthe client side of a distributed speech recognition scheme,rndue to the reduced number of iterations required to reachrnconvergence. Extensive experimentation using the CMUrnSphinx 4 ASR system along with a preliminarily trainedrnspeaker-independent acoustic model for the Italian language,rnin a setting designed for large-vocabulary continuousrnspeech recognition, demonstrates the effectiveness ofrnthe approach even with small amounts of adaptation data.