MAP Based Speaker Adaptation in Very Large Vocabulary Speech Recognition of Czech

Abstract

The paper deals with the problem of efficient adaptation of speechrecognition systems to individual users. The goal is to achieve betterperformance in specific applications where one known speaker isexpected. In our approach we adopt the MAP (Maximum A Posteriori)method for this purpose. The MAP based formulae for the adaptation ofthe HMM (Hidden Markov Model) parameters are described. Severalalternative versions of this method have been implemented andexperimentally verified in two areas, first in the isolated-wordrecognition (IWR) task and later also in the large vocabularycontinuous speech recognition (LVCSR) system, both developed for theCzech language. The results show that the word error rate (WER) can bereduced by more than 20% for a speaker who provides tens of words (incase of IWR) or tens of sentences (in case of LVCSR) for theadaptation. Recently, we have used the described methods in the designof two practical applications: voice dictation to a PC and automatictranscription of radio and TV news.