An octave-scale multiband spectral subtraction noise reduction method for speech enhancement

A new Noise Subtraction Factor (NSF) is proposed in this paper for Multi-Band Spectral Subtraction (MBSS) method. The proposed NSF is a non-linear function of the band Signal-to-Noise Ratio (SNR) for each pre-defined octave band. This NSF adjusts each acoustic band to compensate human auditory perception. Simulation results based on NOIZEUS database demonstrate that the proposed NSF for the MBSS method outperforms the traditional MBSS technique, especially at low input SNR levels. The residual noise in the enhanced speech signal remains minimum while the speech distortion is less perceived.