Acoustic signal-based leak detection methods play a significant role in manual inspection and are gradually evolving into emerging technologies for remote monitoring. However, conventional deep learning approaches for leak detection from acoustic signals often suffer from complex preprocessing procedures and interference from environmental noise, limiting their generalization performance in other water supply networks. Moreover, the accuracy of leak detection still needs further improvement. To address these challenges, this study first constructs linear spectrograms and log-Mel spectrograms with high time and frequency resolution based on pipeline acoustic signal. These spectrograms are designed to capture both high-and low-frequency features, emphasize short-term dynamic changes and subtle spectral components, and are fed in parallel into a convolutional neural network. A parallel time-frequency attention convolutional block is then introduced for feature extraction, enhancing the network’s ability to capture fine-grained temporal and spectral characteristics. Experiments conducted on real-world pipeline acoustic data and physical simulation datasets demonstrate that the proposed method significantly improves the detection rate of leak events and exhibits strong robustness and generalization capabilities.