Intelligent background sound event detection and classification based on WOLA spectral analysis in hearing devices
* Presenting author
Audio signals of real-life hearing devices typically contain background noises. We aim to build a system model which can automatically separate background noise from noisy speech, and then classify background sound into predefined event categories. In this paper, we propose to use weighted overlap-add algorithm (WOLA) and deep neural networks(DNN) for sound event detection, from recordings of 11 realistic background noise environments(cafe, station, …) combined with human speech at different SNR levels. In our approach, a sound trough detection algorithm is used to retrieve important background sound information. We use WOLA as an advanced algorithm which extracts spectral features by transforming a tiny fraction of time domain signal into frequency domain data represented in 22 channels. Moreover, a feed-forward neural network with one hidden layer and one output layer is used to efficiently recognize diverse event feature patterns, and then produce classification decisions based on confidence values. The overall detection accuracy has achieved 95%, while the event ‘hallway’ has the lowest detection rate at 85%. This detection algorithm has the potential for improving noise reduction in hearing devices by applying distinct gain compensation in frequency channels for each particular detected event.