线谱对(LSP)参数是低速语音编码模型的一组非常重要的参数。为了提高编码效率,在压缩感知的框架下提出了基于冗余字典训练的LSP参数压缩感知量化方案。在研究LSP参数在离散傅里叶变换(DFT)域、离散余弦变换(DCT)域稀疏性的基础上,通过K奇异值分解(K-SVD)算法训练得到线谱对参数的冗余字典和稀疏系数。在编码端对从提取好的LSP参数观测得到低维的测量值,并对其量化,在解码端用正交匹配追踪算法对量化后的观测值进行重构得到重构线谱对参数。实验采用重构信噪比衡量了不同稀疏基下重构的效果,采用谱畸变对量化性能进行测试,仿真结果表明,文章采用的基于冗余字典训练算法在线谱对参数的稀疏表示、重构效果和量化效果上都具有更好的性能。 Line spectrum pair (LSP) parameters are a very important set of parameters for the low-speed speech coding model. In order to improve the coding efficiency, a compressive quantization scheme of LSP parameters based on redundant dictionary training is proposed in the framework of compressed sensing. Based on the sparsity of the discrete Fourier transform (DFT) domain and the discrete cosine transform (DCT) domain of the LSP parameters, the redundant dictionary of the line spectrum pair parameters is trained by using the K-SVD (Singular Value Decomposition) Sparse coefficient. At the coding end, low-dimensional measured values ​​are obtained from the extracted LSP parameters, and quantized. The quantized observed values ​​are reconstructed by the orthogonal matching pursuit algorithm at the decoding end to obtain reconstructed line spectrum parameters. The experiment uses the reconstructed signal-to-noise ratio to measure the effect of reconstruction under different sparse basis. The spectral distortion is used to test the quantization performance. The simulation results show that the sparse representation and reconstruction of on-line spectrum based on redundant dictionary training algorithm Effects and quantification have better performance.