论文部分内容阅读
基于 HMM 的汉语语音识别中,易混淆语音的识别率仍然不高.在分析 HMM 固有缺陷的基础上,本文提出一种使用 SVM 在 HMM 系统上进行二次识别来提高易混淆语音识别率的方法.通过引入置信度估计环节,提高系统性能和效率.通过充分利用 Viterbi 解码获得的信息来构造新的分类特征,从而解决标准 SVM 难以处理可变长数据的问题.详细探讨这种两级识别结构中置信度估计、分类特征提取和 SVM 识别器构造等问题.语音识别实验的结果显示,与采用 HMM/SVM 混合结构的模型相比,本文方法在对识别速度影响很小的情况下可以使识别率有明显提高.这表明所提出的具有置信估计环节的 HMM/SVM 两级结构用于易混淆语音识别是可行的.
HMM-based Chinese speech recognition, the recognition rate of easy to confuse speech is still not high.On the basis of analyzing the inherent defects of HMM, this paper presents a method of using SVM for secondary recognition in HMM system to improve the confusion speech recognition rate By introducing the confidence estimation to improve the system performance and efficiency, the new classification features are constructed by making full use of the information obtained from the Viterbi decoding to solve the problem that the standard SVM can not handle the variable-length data.This paper discusses in detail the two-level identification structure Confidence estimation, classification feature extraction and SVM recognizer construction, etc. The results of speech recognition experiments show that compared with the HMM / SVM mixed model, the proposed method can make the recognition of the recognition speed very small This shows that the proposed HMM / SVM two-level structure with confidence estimation is feasible for confusion speech recognition.