Discriminative Bayesian Classification for Large Vocabulary Continuous Speech Recognition

• Main Authors: Yu-Chien Weng, 翁毓謙
• Other Authors: Jen-Tzung Chien
• Format: Others
• Language: zh-TW
• Published: 2005
• Online Access: http://ndltd.ncl.edu.tw/handle/29504570124506789553

碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 93 === In this thesis, we deal with the issues of model uncertainty and model discriminability when building Bayesian classification rule for large vocabulary continuous speech recognition. In conventional Bayeisan classification, we optimize the criterion of minimum Bayes risk (MBR) where the zero-one loss function is considered. The resulting maximum a posteriori (MAP) classification rule has been applied in many speech recognition systems. To improve discriminability of pattern classifier, it is important to design a discriminative loss function where input speech classified to different models should be properly penalized. In this study, we develop a Bayes factor based loss function. This loss/penalty function is established by performing hypothesis test of input speech corresponding to a target model against a competing model. The predictive distributions of target and competing models are computed to determine Bayes factors. In general, the new classification rule is discriminative and robust since the competing model and parameter uncertainty are considered in loss function. We also realize the proposed discriminative Bayesian classification in word graph based search algorithm. From the estimated word candidates and corresponding states, we can calculate loss functions and used them for word graph rescoring of individual word candidates. In the evaluation of broadcast news transcription using MATBN database, we show the superiority of proposed classification compared to MAP classification.