Objective speech quality assessment tools are vital for evaluating speech processing devices and algorithms. It is well-known that, most of the existing objective speech quality assessment tools, such as perceptual evaluation of speech quality (PESQ), segmental signal-to-noise ratio (segSNR), and log-spectral distance (LSD), rely on the knowledge of the clean speech, which is not always available in most cases. Recently, speech quality assessment without the clean speech, which is denoted as non-intrusive speech quality assessment here, has attracted many researchers' interests. In this paper, we study to predict the PESQ score without the clean speech as the reference signal by using Deep Neural Network (DNN). Both regression and classification methods are considered in training the DNN model. In the regression training, all frame-based features extracted from one sentence are mapped to one PESQ score, which is calculated within this sentence. In the classification training, the PESQ score ranging from -0.5 to 4.5 is equally divided into 100 subintervals, and then all frame-based features extracted from one sentence are mapped to one part of these subintervals. To get the training dataset, the TIMIT corpus are chosen as the clean speech signals, which are mixed with different types of noise at different level of SNR to cover all the range of PESQ score. Experimental results show that, in the testing phase, the DNN model can give a predicted PESQ sore without the clean speech, which is highly correlated with the PESQ score having the clean speech as the reference signal. Subjective listening test results further validate the effectiveness of the proposed non-intrusive speech quality assessment.
