An Improved SVD-Based Approach for Speech Enhancement

In this paper, an improved SVD-based approach for speech enhancement is proposed. What can be gained by the improved approach includes a better performance and higher efficiency. Based on SVD algorithm, clean signal is extracted from the noisy speech by retaining signal components and inhibiting noise components recognized by their singular value and signal-to-noise ratio. But different from the previous, the process of filtering singular values is replaced with a selecting and filtering process controlled by rank and gain. This approach results in a cleaner output signal while keeping the original speech signal from distortion. Methods of reducing the total computation are also considered in this paper, making this approach more efficient, providing more options for real-time speech enhancement and hardware implementation. A prewhitening process is built in the approach and makes it competent in enhancing speech corrupted by both additive white noise and colored noise.