一种基于高斯混合模型的导谱频率参数量化算法

Quantization Approach of ISF Based on GMM

提出了一种高效的基于高斯混合模型(GMM)的导谱频率(ISF)参数量化算法,算法的基本思想是利用高斯混合模型将导谱频率(ISF)参数发送给M个高斯簇,然后由高斯格型矢量量化器来量化相应高斯簇的导谱频率(ISF)参数,最终可以在M个量化值中选出频谱失真值最小的一个作为输出值。在设计高斯格型矢量量化器时,基于率失真理论提出了一种最佳比特分配算法。实验结果显示导谱频率(ISF)参数可以透明地压缩到42 bit/帧,与AMR-WB(G.722.2)的多级分裂矢量量化算法相比,节省了3 bit,减少了55%的存储空间。

An efficient Immittance Spectral Frequency （ISF） parameters quantization algorithm is proposed based on the Gaussian mixture model （GMM）. The basic idea of the algorithm is the use of GMM to send the ISF parameters into M Gaussian clusters, and ISF parameters are quantized by a Gaussian lattice vector quantizer corresponding to that Gaussian clustering, and the minimal spectral distortion value among the M quantized values is selected at last. In the design of Gaussian lattice vector quantizer, the optimal bit allocation algorithm is proposed based on the rate-distortion theory. The results show that the ISF parameters could be transparently quantized at 42 bit/frame, which saves 3 bit and reduce 58% of the storage compared with the Split-Multi-Stage Vector Quantization （S-MSVQ） algorithm of AMR-WB（ G. 722.2）.