Howling Noise Cancellation in Time–Frequency Domain by Deep Neural Networks

Abstract

With the wide applications of sound reinforcement system, howling has become a major problem affecting system performance due to the acoustic coupling between the speaker system and the microphone when there exists a positive feedback loop. To suppress the howling noise, in recent years, researchers have proposed many acoustic feedback control methods such as frequency shift method, notch filtering, and adaptive feedback cancellation method. However, current methods mainly involve using adaptive filters in either time or frequency domain, which can suppress howling to some extent but may lead to sound distortion, or have limited suppression ability. In this paper, we propose a novel method to suppress howling noise from speech signal by training deep neural networks (DNN) as an adaptive filter in time–frequency domain, where short-time Fourier transform (STFT) is performed to convert the signal from the time domain to time–frequency domain, and to extract complex values as signal features, so that a supervised end-to-end DNN is constructed which can nonlinearly map the complex values of the howling speech to the complex values of the clean speech, aiming for cancelling the howling noise from the feedback signals. Experimental results have demonstrated that the proposed method can suppress the howling noise effectively, and at the same time greatly improve the quality and intelligibility of the processed speech.