Improvement of N-gram language models using accent phrase boundaries

摘要

Current continuous speech recognition systems make much use of segmental features but little use of prosodic features. This paper proposes a novel method to integrate prosodic boundary information into N-gram-based language modeling. In this method, two types of language sub-models are built. One characterizes word transitions crossing accent phrase boundaries and the other not crossing the boundaries. To realize these two sub-models directly from a speech corpus, its size should be comparable to a text corpus used for N-gram model training. However, the preparation of such a large speech corpus is not realistic. To solve this problem, we focus upon transition of words in terms of their part-of-speech (POS), and differences in FOS transition crossing and not crossing the boundaries are used to generate the two sub-models. Through experiments, the proposed model showed 11% perplexity reduction given the correct boundary position, and 8% reduction with the automatically extracted boundaries. Even when test speech samples were spoken by another speaker than the speaker used in characterizing the POS transitions, 6% reduction was observed.