Abstract
Most of the current computer assisted pronunciation training (CAPT) systems apply the concept of pronunciation confusion networks (PCN) for reliably detecting erroneous pronunciations. However, usually the PCN is constructed manually by subjective rules without considering speaker-dependent pronunciation variations. This thesis proposes an approach to alleviate this problem. Firstly, with the mapping between phone segments of the L2 and acoustic models of the mother tongue, the L1-assisted bilingual PCN is incorporated into our CAPT system for detecting errors of accented non-native speech. Secondly, speaker-dependent pronunciation errors are collected by an iterative procedure automatically to construct a personal PCN. Both types of PCNs take L1 and an individual into consideration, trying to construct the overall recognition network for pronunciation training in a flexible and practical way. Our experiments show that the PPCN yields the best improvement, though the performance of BPCN downgrades when comparing with the PCN baseline system. The possible reason is that the BPCN generated much more speaker-independent paths for individuals, and even worse providing higher perplexity. By using PPCN, the detection rate of erroneous phones of accented non-native speech is increased from 77.6% to 81.57%, when compared to the manually constructed PCN.