矢量量化中线谱对参数失真度量准则的确定

The Determination of the Distortion Measure Using in LSP Parameter's Vector Quantization

线谱对是一种频域的线性预测参数,在如何对其进行有效地量化方面,人们已经进行了广泛的研究。本文从人的主观感觉出发,将加权均方误差准则应用于线谱对参数的矢量量化中,并用自激励编码方式来检验量化后合成语音质量。实验结果表明,在感觉上8bits线谱对参数的矢量量化与34bits线谱对参数的标量量化几乎产生相同的合成语音质量。

Line spectrum pair (LSP) is a kind of frequency domain linear predective parameter. In order to quantize LSP efficiently, people have done a widely study. In this Paper, from the point of human perception, we introduce the weighted mean square error into the vector quantization of the LSP parameter. At the same time, self- excitation vocoder is utilized to verify the synthesized speech quality. The experiment results show that, in perception, the synthesized speech quality with 8bits vector quantization of LSP parameter is very close to that with 34bits scalar quan-tization of LSP parameter.