On the robustness of parametric watermarking of speech
MCAM'07 Proceedings of the 2007 international conference on Multimedia content analysis and mining
| Hi-index | 0.01 |
Parameter-embedded watermarking of speech is effected through slight perturbations of parametric models of deeply-integrated dynamics of the signal. This research focusses on speech watermarking techniques based on linear-in-parameters speech models. Information is embedded by modifying the linear predictor coefficients of the original speech, subject to fidelity constraints. The modified parameters are used to reconstruct the watermarked speech. Experiments with real speech data are used to assess robustness and other performance properties. A particular example of watermark detector design is discussed and performance tested. In set-membership filtering (SMF) based parametric watermarking, linear predictor (LP) coefficients of the original speech are modified subject to an objective fidelity constraint. SMF is used to obtain a hyperellipsoidal set of allowable parameter perturbations (i.e., watermarks) subject to a constraint on the error between the watermarked and original material. This research discusses the robustness of SMF based watermarking to filtering, quantization and combination attacks. An important consideration in watermark robustness is the energy of the watermark signal (difference between watermarked and original signals). Watermarks of higher energy are obtained from perturbed LP coefficients at the boundary of the hyperellipsoidal set. A constrained optimization problem is solved to obtain the best watermarks for filtering and quantization attacks. Finally, a generalized framework for parametric speech watermarking is presented. In addition to the LP model, other parametric representations such as log area ratio, inverse sine, line spectrum pair, and reflection coefficients are used for speech watermarking. An application of perturbed parameter theory for autoregressive models is presented. The perturbed parameter theory is used to obtain bounds on the perturbation of the stegosignal caused by watermarking.