Desynchronization attacks proved to be the greatest challenge to audio watermarking systems as they introduce misalignment between the signal carrier and the watermark. This paper proposes a DNN-based speech watermarking system with two adversarial networks jointly trained on a set of desynchronization attacks to embed a randomly generated watermark. The detector neural network is expanded with spatial pyramid pooling layers to be able to handle signals affected by these attacks. A detailed training procedure of the aforementioned DNN system with gradual attack introduction is proposed in order to achieve robustness. Experiments performed on a speech dataset show that the system achieves satisfactory results according to all the benchmarks it was tested against. The system preserves signal quality after watermark embedding. Most importantly, the system achieved resistance to all considered desynchronization attacks. The majority of the attacks cause less than [Formula: see text]% of incorrectly detected watermarked bits on average, which outperforms comparative techniques in this regard.
