Non-intrusive Speech Quality Assessment using Natural Spectrogram Statistics

Speech quality assessment is one of the active research area in the field of communication and signal processing. In this paper, we proposed a new method to predict the quality of non-intrusive speech signals. This work uses the natural spectro-gram statistical (NSS) properties of speech signals. Undistorted speech follows a natural pattern, which is changed in the presence of distortion. The deviation of NSS in the presence of distortion is used to assess the quality of speech signals by extracting features using the generalized Gaussian distribution and mean subtracted contrast normalized coefficients of the spectrogram. The proposed methodology assess the quality of speech signals without the use of reference speech signal. Experimental results show that the proposed methodology gives high correlation of 0.92 and 0.89, and lowest root-mean-squared error of 0.16 and 0.21 on NOIZEUS-930 and CSTR VCTK Corpus datasets respectively when compared with state-of-the-art speech quality assessment techniques.