Speech Noise Reduction via Intelligent Spectral Gain Selection and Modification

Abstract

Speech information is distributed across the frequency band of 300 Hz to about 3000-4000 Hz. Noise affects speech information distinctively between different spectral components of the speech signal. This means that noise infusion to individual spectral components is variable when measured in signal-to-noise ratio. This variation can be picked up and efficiently discerned by the workings of the human brain. The human brain can perform spectral selection and gain modification to address the various effects of noise in a speech signal. When faced with noise, the brain combines top-down and bottom-up processing to improve the signal-to-noise ratio of the speech signal. The brain's ability to select spectra with more information is not utilized in existing or traditional methods. This paper proposes a novel deep neural network-based method for voice noise reduction. The formulations are determined by attention-based neural networks using spectral gain adjustments as the basis for the proposed technique. The intelligent spectral gain selection and modification using attention mechanisms are introduced after speech signal preprocessing. The noisy signal passes through spectral decomposition where each component has assigned weights based on the proposed attention network. The work shows how the addition of spectral gain adjustments affects the suppression of noise in voice signals by an attention-network-based algorithm. The attention network can focus on the spectral components that are most informative by suppressing bands where noise is excessive. Based on the identified weights, the spectral component is attenuated or amplified to extract the most information content possible. The proposed framework obtained an SNR and segmented SNR of 7.1138 and 1.9950 respectively, higher than existing methods.