CROSSLINGUAL ADAPTATION OF SEMI-CONTINUOUS HMMS USING ACOUSTIC SUB-SIMPLEX PROJECTION

With the demand on providing automatic speech recognition (ASR) systems for many markets the question of porting an ASR system to a new language is of practical interest. Transferring already existing hidden Markov models (HMM) from a source to the target language is seen as a key step to cope with this task. Typically, such a crosslingual model adaptation task consists of a three step procedure. It starts by polyphone decision tree specialisation (PDTS), specialising the phonetic-acoustic decision tree of the source models to the target language. In a second step initial target language models are predicted out of the adjusted decision tree. Finally, the predicted acoustic models are adapted to the target language using a limited amount of target data. In this work we focus on the final model adaptation step in the case of a system architecture employing semi-continuous HMMs (SCHMM). In contrast to continuous density HMMs (CDHMM), adaptation techniques for SCHMMs are not as well developed. In particular, no powerful transformation based adaptation method for adjusting the information bearing mixture weights of the common prototype densities is on-hand. To overcome this problem we introduce a novel adaptation scheme for SCHMM. The method relies on the projection of retrained model parameters to a solution sub-simplex which is obtained through acoustic regression classes derived from the decision tree of the source models. The performance of the procedure is demonstrated by the transfer of multilingual Spanish-English-German models to Slovenian and to French. In the full paper, reference results for a standard maximum likelihood linear regression (MLLR) approach are given too. 1