基于动态字典学习的欠定盲语音重构算法

Underdetermined blind speech reconstruction algorithm based on dynamic dictionary learning

为提高传统字典学习方法选用固定的语音分段长度重构源信号的精度,提出基于动态字典学习的欠定盲语音重构算法,以提取信号中最优的稀疏表示特征。在欠定语音盲分离的两步法框架下,利用正则化Sim CO字典学习对信号进行稀疏表示,依据最速下降思想通过改变语音分段长度迭代优化信号的重构结果直至收敛,得到信号恢复的总体最优解。实验结果表明,相较传统算法,动态Sim CO字典学习算法进一步提取了信号在字典稀疏域的语音特征,在保证运行成本低的同时有效提高了欠定盲语音的重构质量。

To improve the accuracy of the traditional dictionary learning method for source reconstruction by using fixed speech block length,an underdetermined blind speech reconstruction algorithm based on dynamic dictionary learning was proposed to extract the optimal sparse representation features from the signals.Under the framework of the two-step method for the underdetermined blind source separation,the signals were sparsely represented by regularized simultaneous codeword optimization(SimCO)dictionary learning,according to the idea of the steepest descent,the results of signal reconstruction were iteratively optimized until convergence achieved by changing the block length of speech,so as to obtain the overall optimal solution of signal recovery.Experimental results show that,compared with traditional algorithms,the dynamic SimCO dictionary learning algorithm further extracts the speech features of signals in the dictionary sparse domain,and effectively improves the reconstruction quality of underdetermined blind speech while ensuring low operating cost.