,TASK Independent Utterance Verification Using Garbage Model Based on Hierarchical Phoneme

• 作 者: LIU Jia ZHONG Lin LIU Jing LIU Runsheng
• 机 构: Department of Electronic Engineering, Tsinghua University, Beijing 100084, China
• 出 处: 中国电子杂志（英文版）
• 发表日期: 2001年4期
• 关键词: Utterance verification Confidence measures and rejection Speech recognition.

In this paper, an on-line garbage (OLG) model with the hierarchical phonemes is proposed, which is based on the phoneme levels. Thisnew OLG model achieves much better performance and efficiency to reject incorrect recognition and out of vocabulary (OOV) utterances than the OLG model based on the frame levels. A keyword spotting system and an information checking system based on the speech recognition are used to evaluate the performance of utterance verification. The speech databases with the different nature utterances or the noises are used in the evaluations. Instead of using the whole utterance for the estimate of confidence measure, only the hypothesized part, i.e. the part recognized as the keywords, in the utterance is used. The different estimate methods of the OLG models are studied and compared for the task independent utterance verification. The methods of optimizing the competing setin estimating OLG likelihood are presented in the paper. With 5.4% of the within-domain utterances rejected (FRR=5.4%), the recognition accuracy of the keywords for the telephone speech recognition system is over 95% and at the same time the recognition system can reject 99.5% of the irregular utterances (FAR ≤ 0.5%) and 97.4% of the irrelevant sentences(FAR ≤2.6%).