Selective magnitude subtraction for noise reduction

A new method of spectral subtraction is proposed in this paper. Through selective application of common spectral magnitude subtraction, distortion in the processed speech can be reduced significantly without compromising on the amount of residual noise. The decision for the selection can be obtained from the noisy speech itself. A scheme for reducing the amount of residual noise as well as its musical nature is also included. The proposed algorithm is tested on both white noise and F16 noise with highly favorable results.