Acoustic Feedback Detection for Online Video Conferencing

A method of detecting acoustic feedback in online video conferencing system is proposed. Acoustic feedback often occurs when two or more devices in a room attend online video conferencing and their microphones are on. It results annoying sound, thus it is necessary to detect the sound with high accuracy and low latency. Generally, traditional works are based on the method of thresholding features relating to energy of periodic signal. Contrary to the traditional methods, we used training-test structure with neural network. For this, dataset of over 8 hours is collected. In the method, acoustic input of 64ms is classified as feedback frame or ordinary frame. In our experiments, the frame-level classification accuracy of 89.30% is achieved.