Enhancement of Speech Signals Using Multiple Statistical Models

Abstract

This paper proposed a correlative analysis of two recent approaches for noise reduction based on minimum mean-square error (MMSE) and hard thresholding or shrinkage for a provided speech signal. Based on designed algorithm, both noise reduction techniques are real-time data driven schemes. The corrupted speech signal is disintegrate into vacillate frames by a well know sifting process. The fundamental principle of these two techniques is the reconstruction of signal after removal of the noisy signal frames that are already filtered, applying the Minimum Mean-Square Error (MMSE) filtering technique, and hard threshold utilizing shrinkage. These techniques are investigated and correlated with standard denoising methods using different evaluation methods such as, signal-to-noise ratio (SNR) and signal-to-distortion ratio (SDR). The literature of these schemes is retrained to noisy signals that are corrupted by white noise. Experimental results attained demonstrate that proposed MMSE filtering scheme for denoising perform better than hard thresholding or shrinkage in both the segmental and overall consideration of noisy corrupted speech signals.