基于Sparse K-SVD学习字典的语音增强方法

Speech Enhancement Based on Sparse K-SVD Dictionary Learning

摘要：提出一种基于SparseK-SVD学习字典的语音增强方法，采用SparseK-SVD算法自适应地训练一个可稀疏表示的冗余字典，在该冗余字典上采用正交匹配追踪（OMP）算法对带噪语音信号进行稀疏分解，利用稀疏系数矩阵重构纯净语音，实现语音增强．使用NOIZEUS语音库进行了一系列的语音增强实验，主客观评测数据表明，基于稀疏表示的语音增强方法（分别使用SparseKSVD和K-SVD训练字典）相对于传统语音增强方法（小阈值波法、谱减法、改进谱减法）可进一步改善语音质量；对字典训练时间进行统计，发现SparseK-SVD算法训练字典消耗的时间为K-SVD算法训练时间的1／6～1／10，大幅度提高了计算效率．

A speech enhancement method based on sparse representation with Sparse K-SVD dictionary learning algorithm is proposed.Sparse K-SVD algorithm was employed to train a redundant dictionary,which could be sparsely coded on a basis dictionary and learned with only a few sparse coefficients updated.The speech signal′s sparsest coefficients were decomposed through orthogonal matching pursuit algorithm.The speech signal was reconstructed and speech enhancement was achieved.A series of experiments were conducted on the NOIZEUS speech corpus.And the results show that,speech enhancement methods based on sparse representation(using Sparse K-SVD and K-SVD dictionary learning algorithms)outperform the traditional methods(such as wavelets,spectral subtraction,optimized spectral subtraction)both in subjective and objective evaluations.Furthermore,in terms with dictionary training time,the method exploiting Sparse K-SVD performs more efficient than the method using K-SVD by reducing nearly 80%-90%computational complexity.