ID 原文 译文
39276 实验结果表明，在非线性回声和非平稳噪声的环境下，该算法均取得了较好的回声和噪声抑制效果，语音失真较小。 The experimental results show that the proposed algorithm has better echo and noise suppression effects and less speech distortion in the environment of nonlinear echo and non-stationary noise.
39277 为了提升复杂噪声环境下语音增强效果，该文提出了一种基于双层字典学习的单通道语音增强方法。 A single-channel speech enhancement based on jointly constrained double-layer dictionary learning is proposed to improve the quality of speech in the complex noisy environment.
39278 在训练阶段首先采用干净语音和噪声训练初始化特征子字典，然后基于区分性约束和抗混淆约束的优化函数训练双层联合字典， Firstly, the characteristic sub-dictionaries that describe the clean speech and noisy speech are trained. Then, with the new optimization function of discriminative constraints and anti-substitution constraints, a double-layer joint dictionary is trained.
39279 第一层字典表达语音信号和噪声的可区分分量，而第二层字典表达语音信号和噪声的易混淆成分。 The first layer dictionary expresses the separable components of the speech signal and noisy signal, and the second layer expresses easily decomposed components of the speech signal and noisy signal.
39280 在测试阶段含噪语音在双层联合字典上投影得到稀疏系数矩阵，然后重构得到增强后的语音。 In the test phase, the noisy speech is projected on the double-layer joint dictionary to obtain the sparse coefficient matrix, and then the enhanced speech is reconstructed.
39281 该方法利用目标优化函数的约束性减少“交叉投影”现象的发生，降低了信号在联合字典的混淆，从而进一步提升了语音增强的效果。 The constraint of the objective optimization function is used to reduce the occurrence of “cross-projection” phenomenon and the confusion of the signals in the joint dictionary. Furthermore, we can improve the effect of speech enhancement through the double-layer dictionary.
39282 实验结果表明，从语谱图、全局信噪比(SNR)、主观语音质量评估(PESQ)和对数频谱距离(LSD)四个方面评价，相比于基于稀疏约束非负矩阵分解和改进的维纳滤波的语音增强方法，该方法具有更好的性能，能够更有效地去除噪声。 The experimental results show that compared with the speech enhancement methods based on the non-negative matrix factorization with sparsity-regularized constraints and the improved wiener filtering, in four aspects including Spectrogram, Signal to Noise Ratio(SNR), Perceptual Evaluation of Speech Quality(PESQ) and Logarithmic Spectral Distance(LSD), the proposed method has better performance and can remove noise more effectively.
39283 基于深度神经网络的语音分离方法大都在频域上进行训练，并且在训练过程中往往只关注目标语音特征，不考虑干扰语音特征。 Most speech separation methods based on deep neural networks are trained in frequency domain, and in the process of training, they usually only focus on the features of target speech, without considering the features of interference speech.
39284 为此，提出了一种基于生成对抗网络联合训练的语音分离方法。 For this reason, a speech separation method based on cooperative training of generative adversarial network is proposed.
39285 该方法以时域波形作为网络输入，保留了信号时延导致的相位信息。 This method takes the time-domain waveform as the network's input and retains the phase information caused by the signal delay.