动态时间规正与差别子空间相结合变异语音识别的在线训练方法  

Online Training Method of Stressful Speech Recognition Based on Difference Subspace Integrated with Dynamic Time Warping

作　　者：吕成国[1] 韩纪庆[1] 王承发[1] 高文[1] 

机构地区：[1]哈尔滨工业大学计算机科学与技术学院

出　　处：《信号处理》2005年第1期102-105,共4页Journal of Signal Processing

基　　金：国家自然科学基金(60085001)资助项目黑龙江大学青年科学基金博士启动基金资助项目

摘　　要：应力影响下的变异语音(由于说话人受到重力加速度变化而产生)可以用动态时间规正与差别子空间相结合的方法进行识别,但是该方法空间开销很大,而且训练算法极为复杂。针对该方法提出特征矢量替换法和特征矢量求平均法两种基于聚类思想的简单在线训练方法,取一个较小的初始训练集,然后用在线训练的方法使系统识别率达到最大。实验结果表明,其中特征矢量求平均的在线训练方法非常有效,保证系统整体识别率的同时,大大减小了系统的空间开销,适合于训练数据有限的变异语音识别。Stressful speech recognition method based on difference subspace integrated with dynamic time warping could be adopted to recognize speech under G-Force which produced when speaker was under different accelerations of gravity, but the method consumed a mass of memory and the training algorithm was very complex. Two simple online training approaches based on clustering theory to the recognition method, feature vectors substituted method and feature vectors averaged method , were proposed in this paper. A small primary training set could be selected and made the system obtain the best recognition rate by online training. The experimental results showed that the feature vectors averaged online training method was very effective, which reduced system memory consume greatly and ensured the system have good performance at the same time, it was fit for stressful speech recognition when the training data was limited.

关 键 词：变异语音识别 基音频率 动态时间规正 差别子空间 识别率 

分 类 号：TN912.34[电子电信—通信与信息系统]