Complex valued neural networks for audio signal processing

Complex-valued neural networks (CVNNs) were first developed some time ago, but there has recently been renewed interest in extending currently used neural network architectures to enable the use of complex valued data. This paper investigates the benefits of CVNNs compared to conventional real-valued neural networks (RVNNs) for speech enhancement problems. Clean speech signals are mixed with background noise at different signal-to-noise ratios and the networks are then trained to denoise the speech signals in the frequency domain. For the comparison of separation performance, the properties of the complex Ideal Ratio Mask (cIRM) previously proposed are investigated and some preliminary results are discussed with an emphasis on future potential applications