Non-stationary Noise Cancellation Using Deep Autoencoder Based on Adversarial Learning

Abstract

Studies have been conducted to get a clean data from non-stationary noisy signal, which is one of the areas in speech enhancement. Since conventional methods rely on first-order statistics, the effort to eliminate noise using deep learning method is intensive. In the real environment, many types of noises are mixed with the target sound, resulting in difficulty to remove only noises. However, most of previous works modeled a small amount of non-stationary noise, which is hard to be applied in real world. To cope with this problem, we propose a novel deep learning model to enhance the auditory signal with adversarial learning of two types of discriminators. One discriminator learns to distinguish a clean signal from the enhanced one by the generator, and the other is trained to recognize the difference between eliminated noise signal and real noise signal. In other words, the second discriminator learns the waveform of noise. Besides, a novel learning method is proposed to stabilize the unstable adversarial learning process. Compared with the previous works, to verify the performance of the propose model, we use 100 kinds of noise. The experimental results show that the proposed model has better performance than other conventional methods including the state-of-the-art model in removing non-stationary noise. To evaluate the performance of our model, the scale-invariant source-to-noise ratio is used as an objective evaluation metric. The proposed model shows a statistically significant performance of 5.91 compared with other methods in t-test.