Estimation of models for non-native speech in computer-assisted language learning based on linear model combination

This paper investigates how to improve the acoustic modelling of non-native speech. For this purpose we present an adaptation technique to combine hidden Markov models of the source and the target language of a foreign language student. Such model combination requires a mapping of the mean vectors from target to source language. Therefore, three different mapping approaches, based on either phonetic knowledge and/or acoustical distance measures have been tested. The performance of this model combination method and several variations of it has been measured and compared with standard MLLR adaptation. For the baseline model combination small improvements of recognition accuracy compared to the results based on applying MLLR were obtained. Furthermore, slight improvements were found when using an a-priori approach, where the models were combined with predefined weights before applying any of the adaptation techniques.