预测自适应Gauss混合模型线谱频率量化

为了实现高质量低速率语音编码,提出了高效线性预测Gauss混合模型(Gaussianmixturemodel,GMM)线谱频率参数量化算法(LP-GMM-LSFQA)。线谱频率(linearspectralfrequency,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20b/帧时量化性能超过预测分裂矢量量化22b/帧时的量化性能,节约2b/帧;LP-AGMM-LSFQA量化性能优于LP-GMM-LSFQA。

Quantization of LSF parameters based on the linear predictive adaptive Gaussian mixture model

An efficient linear spectrum frequency (LSF) parameter quantization scheme was developed to achieve good reconstruction speech quality in low bit rate speech codecs. In the algorithm, the mean values of the LSF parameters are removed with a first-order inter-frame prediction then used to reduce the inter frame redund-ancy. The residual vectors after prediction are modeled using the Gaussian Mixture Model (GMM) with diagonal covariance matrices and then quantized using random Gaussian distribution vector quan-tization. The means and weights from the GMM are then used in a linear predictive adaptive LSF quantization algorithm. Tests show that the initial algorithm using 20 bit/frame outperforms the linear predictive split vector quantization using 22 bit/frame. The adoptive algorithm gives even better results than the initial algorithm.