Multi-label classification for acoustic bird species detection using transfer learning approach

As part of ornithology, bird species classification is vital to understanding species distribution, habitat requirements and environmental changes that affect bird populations. It is possible for ornithologists to assess the health of a certain habitat by tracking changes in bird species distributions. This work has extended an efficient transfer learning technique for labelling and classifying multiple bird species from real-time audio recordings. For this purpose, Wav2vec is fine-tuned using the back propagation technique, which makes the feature extractor more effective in learning each bird's pitch and other sound characteristics. To perform the task, each audio recording has been clipped as chunks from the overlapping audio to determine multi-labels from it. Through the application of transfer learning, the features of audio recordings have been automatically extracted for classification and fed to a feed-forward network. Subsequently, probabilities associated with each audio segment is aggregated through the clipping approach to represent multiple species of bird call. These probability scores are then used to determine the presence of predominant bird species in the audio recording for multi-labelling. The proposed Wav2vec demonstrates remarkable performance, achieving an F1-score of 0.89 using the Xeno-Canto dataset in which outperforming other multi-label classifiers.