【摘要】 经过传统的联合字典学习算法所训练出来的语音字典会与相同算法下训练出来的噪声字典之间形成相互干扰,残留的噪声以及波形失真都是降噪之后的语音容易产生的问题.针对这一问题,提出一种新的算法,其基于稀疏表示及区分性联合字典学习.为确保信号能够在其对应子字典上进行正确稀疏表示,在字典的学习阶段,这个算法添加了字典区分约束项.最后利用基于区分性联合字典得到的稀疏表示系数对纯净语音进行估计,有效避免了语音失真,获得了更好的语音降噪效果.相比于传统算法,实验结果表明,所提算法在两种评测方式下均获得了最优的评价结果.更多还原

【Abstract】 The speech dictionary trained by the traditional joint dictionary learning algorithm will form mutual interference with the noise dictionary trained by the same algorithm.The residual noise and waveform distortion are all problems that are easily generated after noise reduction.Aiming at this problem,a new algorithm is proposed,which is based on sparse representation and discriminative joint dictionary learning.In order to ensure the signal can be correctly sparse represented on its corresponding sub-dictionary,the dictionary discrimination constraint terms are added to the algorithm during the learning stage of the dictionary.Finally,the sparse representation coefficient based on the discriminative joint dictionary is used to estimate the pure speech,effectively avoiding speech distortion and achieving better speech noise reduction.Compared with the traditional algorithm,the experimental results showed that the proposed algorithm obtained the best evaluation results under the two evaluation methods.更多还原