Effective modeling of acoustic confusions for Mandarin CALL system

Acoustic confusions degrade the accuracy of pronunciation assessment severely in computer assisted language learning (CALL) systems. This paper presents our recent study on effective modeling of the acoustic confusions. We change the traditional Mandarin syllable structure, which is composed of initial and final, to a novel phoneme structure. Several phoneme splitting strategies are investigated, and the question list used for building and merging decision tree is studied. Experiments show that the optimal phoneme splitting strategy outperforms the traditional initial-final structure in our CALL system, with relative 11.05% ASER improvement for nasal finals. This idea may be extended to improve the performance of automatic speech recognition (ASR).