语音识别可信测度和拒识模型新方法研究

This project is a basic research in the applications of speech recognition technology. Confidence measures and rejection models have become important parts in speech recognition systems. In this project, a novel garbage model based on Gaussian mixtures is described; a new utterance verification method based on the on-line garbage model is proposed; the different training methods of the filler model, the competition model, the anti-model and the impostor model are presented; the new algorithms of the supervised and unsupervised speaker adaptation combined with the confidence measures are developed; an method of the confidence estimation of posterior probability based on Multi-Layer Perceptrons (MLP) is proposed; the method of the hierachical averaging and normalized score of the confidence measure based on Chinese pronunciation characteristic are described; a new kind of Viterbi beam searching and pruning algorithm based the confidence measure is proposed; an integrated model based on the multiple confidence information sources is described. Some valuable results are obtained and used for 863 High-Tech project and international cooperation projects. The system robustness and the performance of rejecting noises are improved by using confidence measures in the practical speech recognition systems. . Thirty papers have been published, and one patent applied. One PhD student and 7 MSc have graduated.

语音识别可信测度和拒识模型是口语对话和命令控制系统的关键技术之一。本申请从可信测度估值方法、不同层次结构上可信测度和拒识模型构成及其规一化方法、结合可信测度的说话人自适应方法、可信测度和拒识模型评估方法等方面入手，进行创新性研究，结合汉语特点提出一个完整的新型可信测度和拒识模型算法。该研究具有重要的理论意义和实用价值。