论文部分内容阅读
为了实现高质量低速率的语音编码,提出了高效线性预测Gauss混合模型(Gaussian mixture model,GMM)线谱频率参数量化算法(LP-GMM-LSFQA)。线谱频率(linear spectral frequency,LSF)参数先去均值,经过一阶线性预测,得到残差信号,将残差用协方差矩阵为对角阵GMM量化算法进行量化。在此基础上,利用反量化后参数自适应更新GMM的加权系数和均值,进一步提出了预测自适应GMM-LSF量化算法(LP-AGMM-LSFQA)。实验表明:LP-GMM-LSFQA在20 b/帧时量化性能超过预测分裂矢量量化22 b/帧时的量化性能,节约2b/帧;LP-AGMM-LSFQA量化性能优于LP-GMM-LSFQA。
In order to achieve high-quality and low-rate speech coding, an efficient linear prediction Gaussian mixture model (GMM) spectral parameter quantization algorithm (LP-GMM-LSFQA) is proposed. Linear spectral frequency (LSF) parameters are first averaged, and after the first-order linear prediction, the residual signal is obtained, and the residuals are quantized by using a covariance matrix as a diagonal matrix GMM quantization algorithm. On this basis, the adaptive GMM-LSF quantization algorithm (LP-AGMM-LSFQA) is proposed by adaptively updating the weighted coefficients and the mean of the GMM using the inverse quantization parameters. Experiments show that the quantization performance of LP-GMM-LSFQA is better than that of LP-GMM-LSFQA when the quantization performance of LP-GMM-LSFQA exceeds the quantization performance of 22 b / frame.