Research On Robust Automatic Segmentation Of Dialectal Speech

This paper concentrates on robust automatic speech segmentation of dialectal speech, which is essential for building dialectal speech corpus. Because speech segmenting and labeling are the most time-consuming steps in establishment of speech corpus, by introducing automatic speech segmentation algorithm, we can improve the efficiency of building dialectal speech corpus notably.We carried out systematic study on current automatic speech segmentation algorithms. Based on this, the following aspects of work are done:1. Dialectal speech modeling and its application in automatic speech segmentation.The role of dialectal speech modeling is to adjust model parameters and structure of automatic speech segmentation systems to make sure that the system can be used in dialectal circumstances. In this paper, we compared two classical acoustic model adaptation methods: Maximum a Posteriori (MAP) and Maximum Likelihood Linear Regression (MLLR), which may be supervised by base-form or surface-form phonetic labels. Finally we proposed a dialectal speech modeling scheme which combined multi-pronunciation lexicon and surface-form labels supervised MAP acoustic adaptation.2. A robust framework for automatic speech segmentation of dialectal speechBecause automatic segmentation of dialectal speech is a new research issue, currently there are not effective algorithms proposed by researchers. We're the first one who proposed a framework to deal with this problem. This framework can improve the correction rate of dialectal speech segmentation by 10%.3. Shanghai dialectal Mandarin speech segmentation experimentsWe carried out several segmenting experiments based on speech from Shanghai dialectal Mandarin speech corpus, and the results are reported in this paper.