Linear transformation of Hidden Markov Models based on linear regression

This paper treats a linear transformation of word templates in a word recognition system. The object of the transformation, called LMR-transform, is to adapt the recogniser to a new acoustical environment. The transform is derived by linear regression on pairs of word utterances from two different acoustical environments. The use of the transform has been evaluated for recognition accuracy, speaker-independence and training data requirements in three different fields of application. The main field is noise adaptation in a car-environment. Other fields involve microphone and speaker adaptation. A database recorded in a car incorporating six speakers using five different microphones has been utilised for evaluation. The transform increased recognition accuracy with all of the evaluated fields of application.