A Study on Efficiency Improvement for Discriminative Training Approaches

• Main Authors: Kai-wen Chung, 鍾凱雯
• Other Authors: Bor-shen Lin
• Format: Others
• Language: zh-TW
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/b8v334

碩士 === 國立臺灣科技大學 === 資訊管理系 === 99 === Discriminative training methods have been proven to achieve higher accuracies than the conventional maximum likelihood method for speech recognition system. It can increase the distance between the correct model and competing models by updating competing models in the opposite direction so as to improve the discriminative capability between models. Such training methods require a lot of time and space, which was not yet well discussed in early studies. This paper focuses on the issue of selecting more compact word graph in discriminative training through N-best selection and beam search in order to increase the training efficiency. Experimental results show that (1) Reducing the N-best paths can save 83.7% training time without sacrificing the recognition rate significantly; (2) Beam search can eliminate the less competitive nodes during search, avoid over-training of competing models and improve the accuracy slightly; (3) Minimum phone error (MPE) training can effectively increase the distance between the correct model and competing model and obtain more compact word graph.