Generative Adversarial Network with Guided Generator for Non-stationary Noise Cancelation

Noise comes from a variety of sources in real world, which makes a lot of non-stationary noises, and it is difficult to find target speech from noisy auditory signals. Recently, adversarial learning models get attention for its high performance in the field of noise control, but it has limitation to depend on the one-to-one mapping between the noisy and the target signals, and unstable training process due to the various distributions of noise. In this paper, we propose a novel deep learning model to learn the noise and target speech distributions at the same time for improving the performance of noise cancellation. It is composed of two generators to stabilize the training process and two discriminators to optimize the distributions of noise and target speech, respectively. It helps to compress the distribution over the latent space, because two distributions from the same source are used simultaneously during adversarial learning. For the stable learning, one generator is pre-trained with minimum sample and guides the other generator, so that it can prevent mode collapsing problem by using prior knowledge. Experiments with the noise speech dataset composed of 30 speakers and 90 types of noise are conducted with scale-invariant source-to-noise ratio (SI-SNR) metric. The proposed model shows the enhanced performance of 7.36, which is 2.13 times better than the state-of-the-art model. Additional experiment on −10, −5, 0, 5, and 10 dB of the noise confirms the robustness of the proposed model.