Some Investigations on Attention Mechanism based Deep Learning Models for Speech Enhancement

Abstract

Noise is all around us. When individuals speak, excessive environmental noise newlinecreates transmission issues and has a severe negative impact on intelligibility and speech newlinequality. To address this issue, speech enhancement methods are used to extract clean newlinespeech from environmental disturbances. newlineIn first part, we propose a novel single channel speech enhancement algorithm using newlineiterative constrained Non-negative matrix factorization (NMF) based adaptive Wiener newlinegain for non-stationary noise. The Wiener filter performance depends on the adaptive newlinegain factor value. The adaptive gain factor (and#945;) value is constant regardless of noise type and signal to noise ratio (SNR), so it will affect speech enhancement performance. To overcome this, the adaptive factor value is calculated using a genetic algorithm (GA). newlineHere, the GA adjusts the adaptive Wiener gain based on noise type and SNR level. The newlineGA-based adaptive Wiener gain minimizes Wiener filter estimation errors and improves newlinespeech quality by adjusting the base vector weights of noise and speech. Additionally, newlinethe iterative constraints NMF (IC-NMF) method for calculating the priors from noisy newlinespeech magnitudes. We select the Erlang, Inverse Gamma, Students-t, and Inverse Nak- newlineagami distributions for speech priors and Gaussian distributions for noise priors. Noise and speech samples are well correlated with those distributions. newlineIn the second part, we propose a U-Net with a gated recurrent unit and an efficient newlinechannel attention mechanism for real-time speech enhancement. The proposed U-Net newlinemodel uses skip connections to improve information flow. A novel cross-channel in- newlineteraction can be implemented via the ECA module without dimensionality reduction. newlineIn module testing, choosing an adaptable kernel size for the ECA improved network newlineperformance significantly. Additionally, the U-Net architecture uses gated recurrent newlineunits (GRU), which yields a causal system suitable for real-world use. GRU is used for newlinelearning long-range dependencies. newlineIn the third part, the advanced improv