Speech quality evaluation (SQE) under complex noisy environment is important for audio processing systems and quality of service. Recently, the non-intrusive SQE is getting more and more attentive due to its efficient and ease of use. However, non-intrusive SQEs are expected to be underperformed the intrusive ones since it has no prior knowledge of the clean speech. In this paper, a novel quasi-clean speech reconstruction method for non-intrusive SQE is proposed. The method incorporates Bayesian NMF (BNMF) with deep neural network (DNN), which takes the advantages of both NMF and DNN. BNMF is utilized to calculate the basic spectro-temporal matrixes of target speech, and the obtained matrices are integrated into the DNN model as an individual layer. Then DNN is trained to learn the complex mapping between the target source and the mixture signal, and reconstruct the magnitude spectrograms of the quasi-clean speech. Finally, the reconstructed speech is regarded as the reference of the perceptual model to estimate the Mean opinion score of the tested noisy sample. The experiment results show that the proposed method outperforms the comparative non-intrusive SQE algorithms under challenging conditions in terms of objective measurement.

复杂噪声环境下的语音质量评估（SQE）对于音频处理系统和服务质量至关重要。近来，非侵入式SQE由于其高效且易于使用而变得越来越引人注目。但是，由于非介入式SQE尚不具备干净语音的先验知识，因此预期其性能将不及介入式SQE。提出了一种新的非侵入式SQE准清洁语音重构方法。该方法将贝叶斯NMF（BNMF）与深度神经网络（DNN）结合在一起，从而充分利用了NMF和DNN的优势。BNMF用于计算目标语音的基本时空矩阵，并将获得的矩阵作为单独的层集成到DNN模型中。然后训练DNN学习目标源和混合信号之间的复杂映射，并重建准清晰语音的幅度谱图。最后，将重构后的语音作为感知模型的参考，以估计被测噪声样本的平均意见得分。实验结果表明，在客观测量方面，该方法在挑战性条件下优于非侵入式SQE算法。