Noise Suppression Using a Differential-type Microphone Array and Two-dimensional Amplitude and Phase Spectra

This study aims to achieve noise suppression by processing the output of a microphone array with artificial neural networks (NNs). A differential-type array is used to avoid nonlinear distortions produced by a nonlinear system, such as an NN. The output of the array is considered an image, and it is transformed into a 2-dimensional (2D) spectrum. In the 2D spectrum, the frequency components of a noise are perfectly localized as direct current (DC) components along the spatial frequency axis. In this study, noise suppression was performed by spectral subtraction, after the DC components of noise were instantaneously estimated from the amplitude and phase spectra using independent NNs. As a result, the noise reduction performance of the proposed method with a 16-cmlong array was approximately 24 dB. Although the NNs were trained by white noise, the system was effective for speech and music signals as well as for white noise.