基于Gaussian混合模型的LSF参数量化方法

为了高效率量化线谱频率(linear spectrumfrequency,LSF)参数,提出了基于G auss ian混合模型(G auss ian m ix ture m ode l,GMM)的LSF量化算法。假设LSF矢量属于GMM中的某一个G auss ian分布,用G auss ian分布随机矢量的量化方法对LSF矢量进行了量化。利用准确的G auss ian分布变量量化误差,得到了G auss ian分布矢量的比特分配方法。应用G auss ian分布随机变量的非均匀量化方法量化每一维LSF参数。最后给出了分裂矢量量化、基于概率密度函数(probab ility dens ityfunction,PDF)量化方法和该算法的性能对比。该无记忆LSF量化算法在21 b/帧可以达到透明量化,比传统Sp litVQ节省3 b。

Quantization of LSF parameters using a Gaussian mixture model

An efficient linear spectrum frequency(LSF) parameter quantization scheme was developed based on the Gaussian mixture model(GMM).The algorithm assumes that the LSF parameter has a GMM Gaussian distribution so that the LSF vector can be quantized using a random Gaussian distribution vector quantization.The bits of the LSF parameters are allocated according to the precise quantization error of the Gaussian distribution.Each dimension of the LSF parameter is quantized using a non-uniform scalar quantization of the Gaussian distribution variable.Comparison of the method with the Split-VQ and PDF VQ methods shows that the LSF parameters could be transparent quantized at 21 b/frame by the memoryless quantizer,which is 3b less than the conventional Split-VQ method.