Integrating the Perceptual PMSQE Loss into DNN-based Speech Watermarking

Speech and audio watermarking has been an active research topic during the last thirty years. However, unlike other signal processing techniques, implementations based on deep neural networks (DNN) are relatively recent and many issues remain unexplored. In this paper, we focus on speech watermarking and a key requirement such as the imperceptibility of the watermark. In particular, we explore the application the Perceptual Metric for Speech Quality Evaluation (PMSQE) loss function, originally proposed in the context of speech enhancement, for achieving this goal. In particular, we examine the training trade-offs associated to the watermarking system training procedure and look for a suitable way of incorporating the PMSQE loss. Our experimental results show that the PMSQE loss can, not only meaningfully improve the perceptual quality of the watermarked speech, but also keep, or even improve, other audio quality measures and the bit error rates yielded by attacked signals.