动态时间规正与差别子空间相结合的变异语音识别方法

Stressful speech recognition method based on difference subspace integrated with dynamic time warping

分析了由于说话人受到重力加速度变化而产生的变异语音(应力影响下的变异语音)的特点,把变异语音分成主体部分和变异部分两方面进行研究,提出一种动态时间规正与差别子空间相结合的孤立词语音识别方法。该方法构造差别子空间去除变异部分的影响,利用语音的主体部分进行识别,采用动态时间规正技术对语音特征矢量进行长度对齐,并提出了相应的判别标准。实验结果表明,本方法对应力影响下的变异语音具有良好识别效果,对15个词的小词表,系统平均识别率达到98.3%,与正常语音的识别率基本相当。克服了话者在应力影响下由于心理紧张和生理情况的变化,语音发生严重变异,导致常规语音识别系统性能急剧下降的缺点。

Speech under G-Force was analyzed and considered as principal part and stressful part to research, which produced when speaker was under different acceleration of gravity. An isolated word recognition approach was proposed which integrated difference subspace means with dynamic time warping technique. The method recognized speech under G-Force by constructing a difference subspace to remove the stressful part. Dynamic time warping technique was adopted to make all feature vectors of one word in the training set have equal length, and a corresponding decision criterion was suggested. The experiments showed that for a small vocabulary including 15 words, the method obtained the average recognition rate of 98.3% , which almost equal to the rate in normal environment. The performance of general recognition system was degraded violently for the stressful speech, since G-Force had a direct physical impact on human speech production in addition to the influence on psychology. The method overcame the shortcoming perfectly, not only worked well in normal conditions but also had good performance for speech under G-Force.