Using Chi-Square Testing in Modeling Confusion Characteristics for Robust Phonetic Set Generation

Using Chi-square Testing in Modeling Confusion Characteristics for Robust Phonetic Set Generation Yeou-Jiunn Chen(1) and Chung-Hsien Wu (2) (1) Advanced Technology Center, Computer & Communications Research Laboratories, Industrial Technology Research Institute, Hsinchu, Taiwan, R.O.C. (2) Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan, Taiwan, R.O.C. chenyj@itri.org.tw, chwu@csie.ncku.edu.tw Abstract A phonetic representation of a language is used to describe the corresponding pronunciation and synthesize the acoustic model of any vocabulary. In order to obtain better phonetic representation, context-dependent units are used to model co-articulation effects between phones and have been broadly in speech recognition. However, this representation generally increases the number of recognition units. A phonetic representation with smaller phonetic units such as SAMPA-C for Mandarin Chinese can be applied to reduce the number of recognition units. Nevertheless, smaller phonetic units such as SAMPA-C will contain confusion characters and generally degrade the recognition performance. In this paper, a statistical method based on chi-square testing is used to investigate the confusion characteristics among phonetic units and develop a more reliable phonetic set, named modified SAMPA-C. Finally, experiments on continuous Mandarin telephone speech recognition were conducted. Experimental results show an encouraging improvement on recognition performance can be obtained. In addition, the proposed approaches represent a good compromise between the demands of accurate acoustic modeling. 1. Introduction From the viewpoint of speech recognition, a phonetic representation is functionally defined by the mapping of the fundamental phonetic units of a language to describe the corresponding pronunciation and synthesize the acoustic model of any vocabulary. In the past years, context- dependent units have been broadly used to model the co-articulation effects such as triphone models, which consider both left and right phonemes at the same time. However, this representation generally increases the number of recognition units. Approaches for designing a smaller number of phonetic units are needed in the context-dependent based recognition. In recent years, many phoneme-based phonetic representations have been used such as International Phonetic Alphabet (IPA) [1], Speech Assessment Methods Phonetic Alphabet (SAMPA) [2], and SAMPA for Chinese (SAMPA-C) [3]. Among these representations, SAMPA-C is more flexible and consistent than other phoneme-based phonetic representations for Mandarin Chinese. However, in SAMPA-C, several phonetic units with short duration are not easy to be distinguished and therefore degrade the recognition performance. For Mandarin speech, the confusion characteristics can be found and analyzed in syllable- dependent, subsyllable-dependent, or phoneme-dependent situation. In a training database, syllable- dependent confusion characteristics are difficult to extract due to the sparse data problem. In contrast, the inconsistent phoneme segment in the training data is also not suitable to detect the phone-dependent confusion characteristics. The misdetected phones will result in misrecognition of syllables/subsyllables. Consequently, the phone-dependent confusion characteristic is not helpful for the analysis and representation of confusion characteristics of SAMPA-C based Mandarin speech recognizer. Therefore, the subsyllable is chosen as a compromising unit for the analysis of subsyllable-dependent confusion characteristics. In this paper, based on the statistical hypothesis, the χ2 (chi-square) testing [4] is an alternative test for evaluating …

Content truncated to 500 words