Use of tone information in cantonese lvcsr based on generalized character posterior probability decoding
Tan Lee, Yao Qian · 2005
This thesis addresses two major problems of the existing large vocabulary continuous speech recognition (LVCSR) technology: (1) inadequate exploitation of alternative linguistic and acoustic information; and (2) the mismatch between the decoding (recognition) criterion and the performance evaluation. The study is focused on Cantonese, one of the major Chinese dialects, which is also monosyllabic and tonal. Tone is somewhat indispensable for lexical access and disambiguation of homonyms in Cantonese. However, tone information into Cantonese LVCSR requires effective tone recognition as well as a seamless integration algorithm. Automatic recognition of Cantonese tones has long been regarded as a difficult task. Cantonese has one of the most complicated tone systems among all languages in the world. This thesis presents a novel approach of modeling Cantonese tones. We propose the use of supra-tone models. Each supra-tone unit covers a number of syllables in succession. The supra-tone model characterizes not only the tone contours of individual syllables but also the transitions among them. By including multiple tone contours in one modeling unit, the relative heights of the tones are captured explicitly. This is especially important for the discrimination among the level tones of Cantonese. The decoding in conventional LVCSR systems aims at finding the sentence hypothesis, i.e. the string of words, which has the maximum a posterior (MAP) probability in comparison with other hypotheses. However, in most applications, the recognition performance is measured in terms of error rate (or accuracy). In Chinese languages, given that word is a rather ambiguous concept, speech recognition performance is usually measured in terms of the character error rate. In this thesis, we develop a decoding algorithm that can minimize the character error rate. The algorithm is applied to a reduced search space, e.g. a graph or the N-best sentence list, which results from the 1st pass of search, and the generalized character posterior probability (GCPP) is maximized. (Abstract shortened by UMI.)