Divergence detection in a speech-excited in-service non-intrusive measurement device

摘要

This paper proposes new divergence detection techniques for implementation within in-service non-intrusive measurement devices (INMDs) in public switched telephone networks (PSTNs). The in-service non-intrusive measurement system of interest is used to monitor the delivered quality of speech (QoS) by monitoring the echoes in the telephony network. INMDs are usually based on a class of least mean square (LMS) digital adaptive filters (DAFs). The performance criterion is defined by the modelling convergence rate derived from the optimal Wiener weights, and the excitation for the DAFs is conversational speech. Four types of divergence detectors (DD) are proposed. These are energy divergence detectors (EDD), log energy divergence detectors (LDD), zero crossing divergence detectors (ZDD) and autocorrelation coefficient divergence detectors (ADD). The proposed DDs are based on the detection of voiced/unvoiced/silence periods and as such act as pattern classifiers. Experimental observations have shown that divergence occurs during the low energy unvoiced segments in high-noise environments. The tap-weight coefficients of the DAF are updated with the new value during the voiced segment while the update of the tap-weight coefficients during unvoiced segments of the speech is frozen. This result is then compared with the perfect divergence detector, which employs the Wiener weight theory. The DD techniques reported produce a significant improvement in the system's performance in a noise-impaired environment. Over one second adaptation (8000 samples) the energy divergence detector, the log energy divergence detector, the autocorrelation divergence detector and the zero crossing divergence detector gave model improvements of 16.93 dB, 15.81 db, 12.48 dB and 11.62 dB respectively at echo to noise ratio (e/N) of 0 dB. The proposed DDs compare well with the ideal (nonimplementable) Wiener DD which gives an improvement of 20.92 dB.