Improving Non-native Speech Recognition Performance by Discriminative Training for Language Model in a CALL System
High non-native speech recognition performance is always a challenge for a CALL (Computer Assisted Language Learning) systems using ASR (Automatic Speech Recognition) for second language learning. Conventionally, possible error patterns, based on linguistic knowledge, are added to the ASR grammar network. However, the effectiveness of this approach depends much on the prior linguistic knowledge. In this paper, we design a new scheme for error prediction using two sequential machine learning methods. The first step of the prediction method is aiming at the generality, in which decision tree-based error prediction is adopted in our previous work. The second step of the training is aiming at removing most of the redundant candidates. For the second step, we propose a method based on discriminative training to judge the error candidates that degrade the ASR performance and remove them from the ASR grammar network. An experimental evaluation shows that the proposed method can effectively improve non-native speech recognition performance by drastically reducing the False Alarm rate. Moreover, the smaller WER (Word Error Rate) also confirms the effectiveness of our method.
Hongcui Wang Tatsuya Kawahara Yuguang Wang
Tianjin University, Tianjin, 300072 Kyoto University, Kyoto, 606-8501, Japan
国际会议
2011亚太信号与信息处理协会年度峰会(APSIPAASC 2011)
西安
英文
1-4
2011-10-18（万方平台首次上网日期，不代表论文的发表时间）