Time-delay estimation is an essential part of a wide variety of signal processing applications. This paper follows up on earlier work for time-delay estimation using neural networks. Nonetheless, this work is specialized in the passive acoustic localization of cetaceans. We built a time-delay database from real cetacean vocalizations. Afterwards, we implemented a supervised estimation, based on high-level features and convolutional neural networks. These features are especially designed to deal with the high dimensionality of the cetaceans vocalizations. Finally, we show that our method outperforms traditional approaches when dealing with a realistic dataset which contains large amounts of noise.