Weakly Supervised U-Net with Limited Upsampling for Sound Event Detection

Sound event detection (SED) is the task of finding identities sound events, as well their onset and offset timings from audio recordings. When complete timing information not available in training data, but only are known, SED should be solved by weakly supervised learning. The conventional U-Net with global weighted rank pooling (GWRP) has shown a decent performance, extensive computation demanded. We propose novel limited upsampling (LUU-Net) threshold average (GTAP) to reduce model size, computational overhead. expansion along frequency axis decoder was minimized, so that output map sizes were reduced 40% at convolutional layers 12.5% fully connected without performance degradation. experimental results on mixed dataset DCASE 2018 Tasks 1 2 showed our GTAP about 23% faster achieved 0.644 tagging 0.531 tasks terms F1 scores, while GWRP 0.629 0.492, respectively. major contribution proposed LUU-Net reduction time being maintained or improved. other method, GTAP, further improved provides versatility for various mixing conditions adjusting single hyperparameter.