UL-net: Fusion Spatial and Temporal Features for Bird Voice Detection

Birdsong detection with machine learning (ML) models can provide valuable insights for informed decision making in conservation efforts. Previous methods usually convert audio to spectrogram and utilize convolution to extract spatial features of spectrogram. In this paper, a new framework (UL-net) is proposed to solve the problem of birdsong detection. Our framework integrates U-net, LSTM (Long Short- term Memory) and residual network, and empirically discusses the influence of incorporating timing information into feature representation. Comparison experiments with ResNet-25, CNN and VGG16 were performed on birdVox-70K, a public data set. Ablation experiments were also performed. In several independent experiments, UL-net achieved an average accuracy of 88.49%, higher than the other methods. In comparison with the Baseline method, we find that the LSTM module does improve the accuracy of model prediction, which demonstrates the positive impact of incorporating temporal information into the feature representation.