Attention-Based Spectral Gain Modifications Applied to Speech Noise Suppression

In noisy environments, the brain combines bottom-up and top-down processing to enhance the signal-to-noise ratio of the speech signal. This suggests that biologically inspired algorithms for speech processing to reduce noise can hopefully perform far better than traditional filtering mechanisms in noise suppression. There are several methods inspired by human auditory systems such as Auditory masking models, Perceptual Wavelet Packet Decomposition (PWPD), Non-negative Matrix Factorization (NMF), Artificial Neural Networks (ANNs), and Evolutionary Algorithms. However, the techniques mentioned above do not capitalize on the ability of the brain to select spectra where more information is contained. This study proposes a novel algorithm based on a deep neural network for speech noise suppression. The noisy signal passes through spectral decomposition where each component has assigned weights based on the proposed attention network. The proposed framework obtained an SNR and segmented SNR of 7.1138 and 1.9950 respectively, higher compared to existing methods.