低速率声码器中残差谱的变维矢量量化

Variable Dimension Vector Quantization of Linear Prediction Residual Spectra for Low bit Rate Vocoder

基于一个基音周期内语音线性预测残差（ＬＰＲｌｉｎｅａｒｐｒｅｄｉｃｔｉｏｎｒｅｓｉｄｕａｌ）波形，提出了一种简单而有效的ＬＰＲ谱变维矢量量化（ＶＤＶＱ）方法，即利用具有固定维数的通用码本对ＬＰＲ谱幅度形状进行矢量量化．通用码本和结构化的ＶＱ相组合减少了存储和计算复杂度，产生了高的量化效率，增强了编码语音的感性质量．计算机仿真结果表明，利用该ＶＤＶＱ技术设计的２４ｋｂ／ｓ语音编码器产生了高质量合成语音．

Based on the waveform of linear prediction residual(LPR) within a pitch period,a simple and effective approach of variable dimension vector quantization (VDVQ) of LPR spectrum is proposed.The shape vector of spectral magnitude of LPR is vector quantized by using the universal codebook having fixed dimension.The combination of the universal codebook and structured VQ reduces storage and computational complexity,delivers a high quantization efficiency and enhances the perceptual quality of the coded speech.The results of computer simulation have shown that a speech coder designed by using this VDVQ technique operating at 2400 bps led to a high synthesized speech quality.