混合激励线性预测语音编码标准中线谱频率量化的研究

线谱频率(Line Spectral Frequency,LSF)是线性预测频谱系数(Linear Predication Coefficient,LPC)有效的编码形式。语音线性预测模型中,LPC反映了声道调制的模型,是影响语音听觉感知重要的参数之一。在混合激励线性预测语音编码(Mixed Excitation Linear Prediction,MELP)标准中,对LSF采用4级码本进行分级式矢量量化。首先,为减少其量化冗余度以降低编码速率,本文提出了一种改进的选择算法,生成了一个2级码本替换之。其次,为提高合成语音质量,依据LSF矢量量化的精度与合成语音质量的关系的实验结果,提出根据人耳听觉感知特性进行LSF量化和评价的方法,并予以实验证明。

Studies of the line spectral frequency vector quantization in mixed excitation linear prediction

Line Spectral Frequency(LSF) is one of the most important parameters of the speech perception, it is also an effective expression of Linear Predication Coefficient (LPC) in a linear prediction model, and it reflects the modulation model of the acoustic tube. In standard Mixed Excitation Linear Prediction (MELP), LSF is hierarchically quantized using a 4-level codebook redundantly. In this paper, firstly, an improved codebook choosing algorithm is proposed by generating a 2-level codebook to replace the original 4-level one, so that to reduce the redundancy. Secondly, the method for perceptual quantization and appraisement is suggested to improve the speech quality, which is based on the experimental result that the speech quality is related to the quantization precision of LSF.