Speech enhancement using spectral over-subtraction and residual noise reduction

In this paper, we propose an improved spectral subtraction method for enhancing noise-corrupted speech. This implementation uses over-subtraction method for power spectral subtraction and a modified residual noise reduction algorithm to eliminate musical noise. Performance of the algorithm was compared with previous spectral subtraction methods using both Gaussian white noise and car noise. Proposed method performs substantially better than spectral subtraction because of the absence of any "musical noise" artifacts in the processed speech, therefore it is suitable for speech recognition.