Objective quality measures for perceptual evaluation in digital audio watermarking

In this study, objective quality measures adopted in speech processing are proposed to evaluate the perceptual quality of the watermarked audio signals. Under different audio watermarking techniques such as spread spectrum watermarking, cepstrum domain watermarking, wavelet domain watermarking, echo hiding and histogram-based watermarking, the dissimilarities between the watermarked and host audio signals are separately quantified in terms of several well-developed quality measures, including the signal-to-noise ratio (SNR), the segmental SNR (segSNR), the cepstral distortion (CD), the log-likelihood ratio (LLR), the Itakura-Saito (IS) distortion, the log-area ratio (LAR) and the weighted spectral slope (WSS) measures. For correlation analysis, subjective listening tests and a commercial evaluation tool 'Perceptual Model-Quality Assessment' (PEMO-Q) are also used to grade the differences. Two types of Pearson's correlation coefficients, that is, average and overall correlation coefficients, are calculated to evaluate the performance of quality measures serving as the predictors of perceptual quality. Extensive experiments indicate that there is a good correlation between the investigated quality measures and subjective difference grades. Specifically, the LAR and LLR measures yield the highest average and overall correlation coefficients for most audio watermarking techniques. Moreover, a large decrease in computation time reveals that quality measures run much faster than PEMO-Q. It is therefore concluded that quality measures can be used for objective assessment of the perceptual quality in audio watermarking.