语音信号多维特征参数可视化

Visualization of Multi-dimensional Feature Parameters of Speech Signals

传统的多维可视化技术不能满足语音信号多维特征参数可视化的需求,论文研究了一种新的多维可视化方法,该方法以三维Splatting算法为基础,在其他维度上逐维展开,能较直观地显示多维特征的分布规律。说话人的MFCC特征参数是典型的多维特征参数,应用多维可视化技术分析了说话人识别中经过端点检测后识别率往往略有下降的原因,还进一步证明了统计模式识别的前提条件:训练样本与测试样本的分布要保持一致,否则会明显影响识别性能。

The traditional multi-dimensional visualization technology can not meet the requirements of multi-dimensional feature parameters visualization of speech signals.This paper studies a new multi-dimensional visualization method based on the three-dimensional Splatting algorithm,which is developed on other dimensions.It can display the distribution law of multi-dimensional features more intuitively.The speaker’s MFCC feature parameters are typical multi-dimensional feature parameters.The multi-dimensional visualization technique is used to analyze the reasons why the recognition rate is slightly decreased after endpoint detection in speaker recognition.The preconditions for statistical pattern recognition are also proved that training samples should be consistent with the test sample distribution,otherwise the recognition performance will be significantly affected.