Deep Noise Suppression Maximizing Non-Differentiable PESQ Mediated by a Non-Intrusive PESQNet

Speech enhancement employing deep neural networks (DNNs) for denoising is called noise suppression (DNS). The DNS trained with mean squared error (MSE) losses cannot guarantee good perceptual quality. Perceptual evaluation of speech quality (PESQ) a widely used metric evaluating However, the original PESQ algorithm non-differentiable, therefore, directly be as optimization criterion gradient-based learning. In this work, we propose an end-to-end non-intrusive PESQNet DNN to estimate scores enhanced signal. Thus, by providing reference-free loss, it serves mediator towards training, allowing maximize score We illustrate potential our proposed -mediated training on strong baseline DNS. As further novelty, train and alternatingly keep up-to-date perform well specifically under training. Detailed analysis shows that mediation increases performance about 0.1 points synthetic test data 0.03 DNSMOS real data, compared MSE-based loss. Our method outperforms Interspeech 2021 Challenge 0.2 data. Furthermore, improves same approximated differentiable loss 0.4