最小方差无失真响应浊音谱建模方法研究

讨论了在语音处理中广泛使用的线性预测模型,分析了线性预测模型在建模语音谱包络方面存在缺点,并分析了线性预测模型过高估计浊音谐波频率处能量以及随着阶数增加线性预测谱逐渐恶化的原因。在此基础上,提出了最小方差无失真响应(MVDR)建模方法,并讨论了通过Toeplitz矩阵的Cholesky分解确定MVDR滤波器系数的方法,通过与线性预测方法进行比较,发现最小方差无失真响应滤波器能提供一个更好的原始语音包络。

Modeling of sonant based on minimum variance distortionless response spectrum

This paper discusses the LP model which is widely used in speech processing, analyzes its shortcomings in modeling speech spectrum envelop and the reason why LP model over-estimates the power at sonant harmonic frequency and that why LP spectrum de- teriorates with the increasing order.The Minimum Variance Distortionless Response Model- ing method is then introduced,and how to determine its matrix filter coefficient by means of Toeplitz matrix’s Cholesky decomposition is discussed.Comparing with the LP modeling method,the Minimum Variance Distortionless Response Filter is found to provide better envelop of original speech.