Subspace Approach for Enhancing Speech based on SVD

In this work, we have developed an improved approach for enhancing speech corrupted by additive white Gaussian noise. An efficient denoising approach based on singular value decomposition (SVD) and Savistky-golay filter is proposed to reduce the White Gaussian noise (WGN). In order to filter singular values (SV) that represent original speech signal, we have proposed an efficient threshold algorithm. The whole SV are extracted from Hankel matrices in the overlapping speech window frames. After selecting dominant SV by our efficient threshold, the inverse operation is performed to create the new Hankel matrices and regrouping frames to reconstruct the enhanced signal (original signal), the denoised signal is smoothed using the Savitzky-Golay filter. The proposed approach offers an improved performance of speech enhancement comparing with traditional methods in terms of Perceptual Evaluation of Speech Quality scores measure (PESQ) and segmental SNR (SegSNR). A good yield of the method is observed for the SNR values between −10 and 30 dB.

• speech enhancement
• speech denoising
• additive white Gaussian noise
• singular value decomposition
• Hankel matrices
• singular value thresholding
• Savitzky-Golay filter
• subspace methods
• PESQ
• segmental SNR

• Speech and Audio Processing
• Image and Signal Denoising Methods
• Image Enhancement Techniques
• speech enhancement
• speech denoising
• additive white Gaussian noise
• singular value decomposition
• Hankel matrices
• singular value thresholding
• Savitzky-Golay filter
• subspace methods
• Signal Processing
• Computer Vision and Pattern Recognition
• Computer Science