Single-Ended Quality Measurement of Noise Suppressed Speech Based on Kullback-Leibler Distances

In this paper, a single-ended quality measurement algorithm for noise suppressed speech is described. The proposed algorithm computes fast approximations of KullbackLeibler distances between Gaussian mixture (GM) reference models of clean, noise corrupted, and noise suppressed speech and a GM model trained online on the test speech signal. The distances, together with a spectral flatness measure, are mapped to an estimated quality score via a support vector regressor. Experimental results show that substantial improvement in performance and complexity can be attained, relative to the current state-of-art single-ended ITU-T P.563 algorithm. Due to its modular architecture, the proposed algorithm can be easily configured to also perform signal distortion and background intrusiveness measurement, a functionality not available with current standard algorithms.