On improving non-stationary noise suppression in telephony using deep neural networks

Today personal audio devices are usually used during telephone connections. Mobility and facility take phone callers into challenging noise environments that challenge speech intelligibility and quality. Noise suppression is an essential sector in speech enhancement during telephone connection. Noise suppression has been a research topic for decades, but traditional noise suppression methods have limits. Traditional noise suppression methods commonly perform well in stationary noise environments where background noise does not change rapidly. Today we want more from noise suppression to work satisfyingly in challenging and rapidly changing environments. Solutions for challenging, non-stationary, and rabidly changing noise environments are searched from deep neural networks. This thesis researches a convolutional neural network as a noise suppression algorithm. The chosen network is a modified version of a network called U-Net. We prestudy U-Net with different loss functions, selecting a good option with the help of objective metrics. After preselection, we chose SI-SDR as our loss function. In this thesis, we arranged a listening test where participants compared blindly traditional noise suppression method to DNN-based noise suppression. Participants of the listening test evaluated the following attributes: speech intelligibility, speech quality, noise transparency, and noise level vs. speech. Nevertheless, of assumptions, U-Net did not outperform in all attributes in non-stationary environments. U-Net increased the speech intelligibility, and noise levels were better compared to speech than with the traditional method. The output of the traditional method was better with speech quality and noise transparency. In conclusion, the DNN-based noise suppression method increases speech intelligibility, but it does not guarantee quality