Neural network based system for optimal watermark embodiment in audio signals

Summary form only given. In this work we present a novel technique for optimal embodiment of watermark, i.e. auxiliary digital signals, into audio host signals based on a properly trained neural network. One of the important practical applications of such a class of systems pertains to automatic royalty tracking and proof of copyrighted material or commercial advertisements, when advertisers are able to confirm that commercials which they have paid for were actually broadcast at the proper time and date. Our method includes introduction of a local detection subsystem on the embedder side. The embedder simulates the extractor process at receiver side, including expected channel distortion, giving reliable estimate of actual watermark detection error rate. Collecting pairs of selected characteristics of host signals and the estimated error rate, it is possible to train a corresponding multilayer neural network which can find the optimal watermark embodiment during the working regime of the embedder.