Abstract
Generally, most blind signal separation algorithms deal with the separation problem in the absence of noise. The presence of noise degrades the performance of separated signals. This paper deals with the problem of blind separation of audio signals from noisy mixtures. Blind signal separation algorithm is applied on the discrete cosine transform, the discrete sine transform or the discrete wavelet transform of the mixed signals, instead of performing the separation on the mixtures in the time domain. All of these transforms have an energy compaction property, which concentrates most of the signal energy in a few coefficients in the transform domain, leaving most of the transform-domain coefficients close to zero. As a result, the separation is performed on a few coefficients in the transform domain. Another advantage of signal separation in transform domains is that the effect of noise on the signals in the transform domains is smaller than that in the time domain. The paper presents also an investigation of the rule of the speech enhancement techniques as pre- and post-processing steps for the blind signal separation process, instead of performing the separation on the mixtures in the time domain. The considered speech enhancement techniques are the spectral subtraction, the Wiener filtering, the adaptive Wiener filtering, and the wavelet denoising techniques. Both blind signal separation and noise reduction are applied within a real speaker identification system to reduce the effect of interference and noise on the system performance. The simulation results confirm the superiority of transform domain separation to time domain separation and the importance of the wavelet denoising technique, when used as a pre-processing step for noise reduction. Moreover, the speaker identification system performance is enhanced with blind signal separation and noise reduction.