Iterative noise power subtraction technique for improved speech quality

An iterative spectral subtraction technique for speech enhancement that gives improved performance has been presented. In all the speech enhancement methods, a predetermined average power is subtracted in the spectral domain or a gain function less than unity is generated for the individual harmonics to reduce noise power. What ever the process is, it is a common observation that some of the power spectrums become negative after the enhancement process and is avoided by forcing the coefficients to zero. Under such a condition the actual amount of noise power subtracted is less than the amount intended. If the power subtraction is done in an iterative way such that the predetermined power is correctly subtracted from the noisy signal, the segmental SNR improves appreciably. The results show that an improvement in average segmental SNR by more than 0.5 dB can be achieved by adopting this technique.