Real-Time Noise Suppression Method Using Deep Learning Algorithm

The capacity of deep learning-based speech enhancement algorithms to effectively eliminate both stationary and non-stationary noise components from noisy speech observations has been demonstrated. However, they frequently add false residual noise, particularly when the training goal lacks phase information, such as an ideal ratio mask or the magnitude and fluctuations of clean speech. It is widely known that the perception speech quality may deteriorate whenever the power of the residual noise components surpasses the noise masking threshold of the human auditory system. One logical approach is to use a post processing strategy to further attenuate the remaining noise components. However, estimating the noise power spectral density (PSD) is a difficult challenge due to the kind of residual noise's very non-stationary character. In order to address this issue, the research suggests three methods for estimating the noise PSD frame by frame. The remaining noise can then be efficiently eliminated by using a gain function developed using a decision-directed methodology.