A Neural Network-based Howling Detection Method for Real-Time Communication Applications

Howling arises from acoustic coupling between the speaker and the microphone when it creates positive feedback. Traditional public addressing systems and hearing aids devices detect and suppress the howling using conventional howling features. However, conventional howling features in real-time communication (RTC) suffer from nonlinearities and uncertainties such as various speaker/microphone responses, multiple nonlinear audio processing, unstable network transmission jitter, acoustic path variations, and environmental influences. In howling detection, the signal processing methods using specific temporal-frequency characteristics are ineffective for RTC scenarios. This paper proposes a convolutional recurrent neural network (CRNN) based method for howling detection in RTC applications, achieving excellent accuracy with low false-alarm rates. A howling dataset was collected and labeled for training purposes using different mobile devices, and the log Mel-spectrum is selected as input features. The proposed method achieves an 89.46% detection rate and only a 0.40% false alarm rate. Furthermore, the model size of the proposed method is only 121kB and has been implemented in a mobile device running in real-time.