基于准KLT域的线谱对参数压缩感知量化研究

Efficient Compressed Sensing Quantization of LSP Parameters Based on the Approximate KLT Domain

用尽可能少的比特数实现线谱对(LSP)参数透明量化一直是语音编码领域的研究热点。该文基于压缩感知理论,研究了LSP参数在准KLT域的稀疏性,并设计了LSP参数先压缩感知再矢量量化的方案。编码端,利用压缩感知理论,在准KLT域将原始LSP参数投影到低维空间,得到低维测量值,而后采用分裂矢量量化算法对测量值进行量化;解码端,以量化后的测量值为已知条件,利用正交匹配追踪算法重构出原始LSP高维矢量,重构值作为最终量化值。实验结果表明,算法在适当的码本存储量和搜索复杂度下,达到透明量化效果所需的比特数最优时仅需5 bit/帧。

For low bit rate speech coding applications,it is very important to quantize the Line Spectrum Pair（LSP） parameters accurately using as few bits as possible without sacrificing the speech quality.In this paper,the sparsity of LSP parameters on the approximated Karhunen-Loeve Transform（KLT） domain is researched,and then an efficient LSP parameters quantization scheme is proposed based on the Compressed Sensing（CS）.In the encoder,the LSP parameters extracted from consecutive speech frames are compressed by CS on the approximate KLT domain to produce a low dimensional measurement vector,the measurements are quantized using the split vector quantizer.In the decoder,according to the quantized measurements,the original LSP vector is reconstructed by the orthogonal matching pursuit method,the reconstructed LSP vector is the ultimate quantization value of the original LSP parameters.Experimental results show that the scheme can obtain transparent quality at 5 bit/frame with realistic codebook storage and search complexity.