잡음제거 모델 훈련을 위한 딥러닝 기반 가상 데이터베이스 생성 기법

Neural network training for deep learning-based noise cancellation requires a large amount of data acquired in a real noise environment, but it is not easy in many respects, such as various costs. In this paper, we propose an alternative method to estimate the environment based on the deep neural network using the proper amount of original sound and noisy signal data. Then, it is possible to generate a large amount of virtual noisy database by inputting the original sound spectrum through the deep neural network and outputting the ideal ratio mask. We can build a large database similar to the actual environment through the proposed method, which can greatly improve the denoising performance.
Experiments in real environments have shown that the proposed method can be used successfully for deep learning-based denoising.