Aplication of spectral subtraction for reducing industrial noises

In this paper, we use spectral subtraction to remove non-stationary industrial noise from speech signal. Motor generated industrial noises often contain a wide-band part and some limited number of dominant sinusoidal signals that they carry most of the noise energy, especially, at low frequencies. The latter components highly drop the speech signal quality. To fulfil the requirements of spectral subtraction, we present a novel approach based on multi-band minimum statistics to estimate the noise. To evaluate the performance of the proposed algorithm, tests are conducted that carry reduction of real angle grinder noise, in both engaged and non engaged phases, from speech samples. Objective and subjective measures reveals that our proposed algorithm can eliminate wide-band angle grinder noise plus the non-stationary dominant frequency components well enough, with minimum speech distortion and musical noise in the processed speech signal.