Perceptually-based objective measures for speech quality assessment

This paper proposes two new perceptually-based objective quality assessment techniques, In these techniques, we emulate several known features of perceptual processing of speech sounds by human ear to map the speech power spectrum (from frequency domain) into auditory power spectrum tin bark domain). The perceptual features derived from auditory power spectrum of input and output coded speech signal are used to estimate the output speech quality, In the first technique, the extracted feature is the average bark spectral distance per band (BSDB) while it is bark spectral distance per word (BSDW) in the second technique, Finally, we use the abductive networks, that evolved from neural network concepts, statistical modeling and artificial intelligent concepts, to estimate the speech quality from BSDB and BSDW in the first and second technique respectively. Results indicate that the first technique is more reliable and robust in evaluating the coded speech quality and it is highly correlated to human responses across a wide range of quality level and for wide range of speech processing.