Multi-label Bird Species Classification Using Hierarchical Attention Framework

The paper uses an attention-based framework to classify bird vocalizations in real-time audio recordings. Transformers have recently demonstrated exemplary performance on classification tasks. Shifted window hierarchical vision transformers (Swin-Transformers) have demonstrated impressive image identification performance. The window shift technique amends effectiveness by constraining self-attention-based computing to local window frames that are non-overlapping while allowing for cross-window connectivity. This architectural hierarchy can simulate at various scales, and the computational complexity grows linearly as the size of the image enlarges. We experimented with mel-spectrogram of bird vocalizations. The analysis is also carried out with a hierarchical attention-BiGRU model over Mel-frequency cepstral coefficient. We followed a short-time sliding aggregation technique to determine the test sample, in which the probability outputs are summed and normalised species-wise. The most probable species are assumed to be birds present in the recording. Our sequential models, attention-BiGRU and Swin-Transformer, perform quite well in the Xeno-Canto database with F1-scores of 0.84 and 0.67, respectively, outperforming state-of-the-art multi-label classifiers.