区分性联合稀疏字典交替优化的语音增强

Speech enhancement based on discriminative joint sparse dictionary alternate optimization

在联合稀疏字典的语音增强中,由于联合字典的相似性,导致稀疏重构阶段产生语音和噪声混淆进而产生语音失真问题。针对此,在训练阶段提出一个费希尔准则下的目标函数。该函数包含了语音和噪声的区分约束项,并用与信号变化相关的平衡因子去调整各项权值,为尽可能减小混淆误差提供了保障;同时,为了能使目标函数收敛,设计了一种交替优化字典和稀疏系数的算法,迭代寻找所需的字典和稀疏系数,完成语音字典和噪声字典的输出,得到具有非相似即区分性能较好的联合字典。在增强阶段,将带噪语音信号在联合字典上进行稀疏表示,并估计出语音幅度谱和噪声幅度谱。最后,结合维纳滤波器和理想二值掩模的优点,提出了新的软掩模滤波器,进一步消除了残余噪声。通过对不同信噪比的带噪语音进行实验,新算法得到的语音信噪比和听觉感知评价都较高,验证了新算法在提高语音性能方面的有效性。

In the speech enhancement of the joint sparse dictionary, due to the similarity of the joint dictionary, the speech and noise confusion is generated in the sparse reconstruction stage, which will generate the speech distortion problem. In view of this, an objective function under the Fisher criterion is proposed in the training stage. This function contains the distinguishing constraint of speech and noise, and adjusts the weights with the balance factor related to the signal change, so as to make the confusion error as small as possible. At the same time, in order to make the objective function converge, an algorithm is designed for alternately optimizing the dictionary and sparse coefficients. The algorithm is iterated to find the needed dictionary and sparse coefficient, and completes the output of the speech dictionary and noise dictionary. A joint dictionary with dissimilarity and good discrimination performance is obtained. In the enhancement phase, the noisy speech signal is represented sparsely in the joint dictionary, and the speech amplitude spectrum and noise amplitude spectrum are estimated. Finally, combining the advantages of the Wiener filter and ideal binary mask, a new soft mask filter is proposed. The residual noise is further eliminated. Through the experiments of noisy speech with different signal-to-noise ratios (SNR), the new algorithm has high SNR and auditory perception evaluation, which verifies the effectiveness of the new algorithm in improving speech performance.