A multi-view cooperative contrastive learning framework for Birdsong Classification

Birdsong recognition plays a vital role in biodiversity research and conservation. Traditional classification models typically require large amounts of labeled data for training, and the process of annotating audio recordings is labor-intensive and time-consuming, posing a significant challenge for large-scale data applications. To mitigate this issue, multi-view contrastive learning (MVCL) provides a solution by extracting informative features from existing data, reducing the number of labels required for downstream tasks. However, most existing MVCL methods focus primarily on intra-view self-supervised learning with single-level feature comparisons, limiting their representational capacity. To address the challenges, this paper proposes a multi-view cooperative-consistency contrastive learning (MVCCL) framework for birdsong recognition. Firstly, three kinds of spectral features are extracted, including wavelet transform (WT) spectrogram, short-time Fourier transform (STFT) spectrogram, and Hilbert–Huang transform (HHT) spectrogram, with each serving as an independent view. Then, instance-level contrastive learning, cross-view cooperative contrastive learning, and cluster-level consistency contrastive learning are designed to strengthen discriminative features within each view, promote cooperative learning across views, and ensure the consistency of multi-view feature representations, respectively. The experiments are conducted on a self-built dataset and a public dataset, Birdsdata. The experimental results show classification accuracies of 97.19% and 96.55%, respectively, outperforming the existing methods. Notably, with only 5% labeled data, MVCCL achieved performances of 75.65% and 70.10%, verifying its effectiveness in small-labeled data scenarios. In conclusion, MVCCL can effectively capture rich cross-view common features and spectrogram details, provide a robust pre-training model, improve birdsong recognition performance, thereby support ecological monitoring and biodiversity conservation efforts. • Designed a high-accuracy birdsong recognition framework with limited labeled samples. • Proposed MVCCL framework with three types of contrastive learning. • Designed cooperative contrastive loss to enhance cross-view information sharing. • Developed consistency loss to reduce variance and align clusters across views. • MVCCL outperformed other methods on public birdsong datasets in comparisons.