Non-intrusive whitening of speech using least mean square and divergence detection technique

A speech whitening technique is presented and used for improved echo path modelling in telephony networks. The system identification of interest is based on the real time least mean square (LMS) algorithm and a class of digital adaptive filters (DAFs). The modelling convergence rate derived from the optimal Wiener weights defines the performance criterion. A novel non-intrusive whitening technique based on the speech characteristics is exploited to whiten the speech power spectral density (PSD) whilst preserving the signal bandwidth requirements. The technique involves pre-filtering the speech using tap weight coefficients of the inverse speech spectrum. Software simulation shows an improved performance compared to the conventional LMS. A new divergence detection (DD) technique is used in a noise-impaired environment to eliminate divergence by controlling the adaptation process. The DD technique reported produces significant performance improvement in noisy environments and at echo to noise ratios (e/N) of up to 0 dB. The combined improvement reported using the whitening technique and DD is 24.5 dB after 8000 iterations (1 second) at e/N of 0 dB.