基于过完备线性预测字典的压缩感知语音重构

基于语音信号帧内样值间的相关性和冗余域的稀疏性,针对采用离散余弦转换矩阵及基追踪方法对压缩感知采样语音进行重构时,语音稀疏性不够好导致大压缩比采样后重构效果差的缺点,提出采用过完备线性预测字典做转换矩阵,用基追踪重构算法对压缩感知采样语音进行高质量重构。该方法预先由训练语音的预测系数聚类构造过完备字典,不需要测试语音的预测系数;基于过完备线性预测字典重构信号性能良好。对利用基追踪重构的语音进行了主客观评价,得出结论:同样的观测数目下,基于过完备线性预测字典比基于离散余弦变换矩阵压缩感知采样语音重构信噪比高出3～8 dB。

Reconstruction of compressed speech sensing based on overcomplete linear prediction dictionary

Using the correlation of speech signal and the sparsity in redundancy,we propose a new algorithm of compressed sensing of speech signal based on overcomplete linear prediction(OLP) dictionary obtained from the codebook constructed from the linear prediction coefficients of training signals.The new method not only improves the performance of reconstructed speech signals based on Gaussian measurement matrix and basis pursuit,but also does not need to know the prediction coefficients of the signals under test.We apply OLP dictionary and discrete cosine transform(DCT) methods to compressed speech sensing and compare these two methods;and the reconstructed speech signals are evaluated with objective and subjective evaluation criteria.Experiment results show that the SNR performance of compressed speech signal sensing based on OLP method is 3-8 dB higher than that using DCT method under the same number of measurements.