基于动态阈值的MVN连续语音特征调整算法

Continuous Speech Feature Adjustment Algorithm Based on Dynamic Threshold MVN

针对非特定人大词汇量连续语音识别,在均值方差归一化的基础上,提出了基于动态阈值的特征调整方法。动态阈值的选取方式包含阈值的动态范围确定和确定阈值的系数。动态阈值范围的确定依据如下两个数值,一个是样本特征点的均值,另一个是使得样本特征点等分的数值。然后再根据对特征点在样本特征点均值上下的比例关系得到系数,最后根据这个系数来确定一个具体的阈值,并基于此阈值对连续语音特征曲线进行调整。

The article focuses on the Speaker independent large vocabulary continuous speech recognition. On the funda-mental of MVN,the proposed speech feature adjustment method based on dynamic threshold. The selection of dynamic threshold includes how to find the range and the coefficient of dynamic threshold. Get the range of dynamic threshold is from the mean of feature vectors and the value, which separates the feature, vectors by an equal division method. Then, we can get the coefficient according to the proportion relationship around the mean of feature vectors. Finally, we compute the threshold by using the coefficient and use the threshold to adjust the continuous speech feature trajectory.