基于压缩感知的K-L分解语音稀疏表示算法

Compressed-Sensing-Based Speech Sparse Representation with K-L Expansion

为克服稀疏基在KLT域不便传输的不足,提出一种基于Karhunen-Loeve(K-L)正交分解的语音稀疏表示算法。结合压缩感知理论,建立语音自相关模型并求解Fredholm积分方程,采用二分法估计出可实时传输的模型参数,构造非相干字典;然后用随机矩阵对语音在字典上的稀疏投影系数进行观测获得低维观测值。重构结果表明:相比已有的稀疏表示算法,本文算法的字典匹配性更好,且具有较好的语音质量。

To overcome the infeasibility of real-time transmission of KLT sparsifying basis, a speech sparse representation algorithm is presented based on Karhunen-Loeve （K-L）expan- sion. With compressed sensing（CS） theory a speech autocorrelation model is built. Fredholm integral equation is solved and the model parameter is estimated. This parameter is used to construct incoherent dictionary and can be easily transferred. Thereafter, low-dimensional measurements are obtained by sensing the sparse vector with a stochastic matrix. Reconstruc- tion experiments show that the proposed algorithm outperforms the existing methods in dic- tionary adaptability and reconstruction oualitv