Comparative study of deep learning based and traditional single-channel noise-reduction algorithms

Deep neural networks (DNN) have been applied to the problem of noise reduction and promising results have been reported widely, leading to the impression that the traditional techniques based on blind noise estimation may no longer be needed. However, there lacks comprehensive and rigorous evaluation and comparison between DNN based and traditional noise-reduction algorithms for their pros and cons. In this work, we attempt to evaluate some widely used DNN based noise-reduction algorithms and compare them to a traditional noise-reduction method. We also evaluate a method that straightforwardly combines a DNN based regression method with the optimal filtering technique. Through experiments, it is observed that: 1) DNN based methods have advantages over the traditional methods in scenarios with non-stationary noise and low signal-to-noise ratios (SNRs); 2) generalization remains a challenging issue with DNN based methods; for noise types unseen in the training data, which happen often in practical environments, DNN based methods do not show any advantage over the traditional technique; 3) combining DNN-based regression and the optimal filtering technique shows some potential in improving the noise-reduction performance as well as system generalization.