Multiple Noise Suppression Based on Simultaneous Learning and Independent Estimation of Amplitude and Phase of Noise
Koki Takenaka, Kenji Ozawa · 2022 International Conference on Engineering and Emerging Technologies (ICEET) · 2022
This study aims to suppress noise by taking advantage of the fact that the target sound components are concentrated in the spatial direct current (DC) bin when microphone array outputs are transformed into a spatiotemporal two-dimensional spectrum. For this purpose, the spatial DC components of the noise spectrum are estimated and subtracted from the observed spectrum in the spatial DC bin to suppress the noise. The amplitude and phase of the spatial DC components of noise are estimated using two independent neural networks (NNs) in which both amplitude and phase information are used simultaneously to consider the mutual dependency between the amplitude and phase. The NNs are learned by assuming the presence of two noise sources. Computer experiments show that approximately 20 dB of noise suppression can be achieved for the cases with one, two, and three noise sources.