Automatic language recognition with tonal and non-tonal language pre-classification

Parallel Phoneme Recognition followed by Language Modelling (PPRLM) systems currently provide state of the art language identification performance on conversational telephone speech. In this paper an innovative method for tonal and non-tonal language pre-classification by using prosodie information is reported. Our motivation is to improve recognition accuracy and save the amount of CPU run-time while handling large number of languages. Also, by incorporating different confidence measures into the traditional PPRLM framework, we propose an optimized language recognition system that can be applied in an open-set language recognition task. For a task of 12 target languages and 4 non-target languages, our results show that with the optimized pre-classification, Universal Background Phone Model confidence measuring and Witten-Bell discounting the system can achieve recognition accuracy rates of 77.9% for 30-sec speech segments and 49.2% for 10-sec speech segments.