Abstract
Traditional automated bird sound monitoring systems typically rely on closed-world classification, which limits their utility when encountering species not present in the training data. This paper proposes a label-agnostic approach using correlation clustering. By training a Siamese neural network to predict whether two audio clips belong to the same species, the system can discover and group sounds from novel species while filtering out environmental noise. Our results show that this method generalizes effectively to unseen data and that the clustering process itself helps refine pairwise predictions.
Introduction
Most current bioacoustic monitoring tools are built on supervised classification models. While effective for a fixed set of species, these models struggle with the "open-world" reality of field recordings, where unknown species or environmental noises are frequently encountered. To address this, we shift the focus from species identification to species discovery through clustering. Instead of asking "What species is this?", our system asks "Do these two sounds belong to the same species?" This allows for the organization of massive, unlabeled datasets and the identification of rare or previously unrecorded vocalizations.
Methodology
The proposed pipeline consists of three primary stages: pairwise learning, probabilistic modeling, and correlation clustering.
Feature Extraction and Pairwise Learning
We utilize a Siamese Neural Network based on the ResNet-18 architecture. The network processes two audio spectrograms through shared weights to extract feature vectors. These vectors are then concatenated and passed through a fully connected network to output the probability that the two clips represent the same species.
Probabilistic Modeling and Correlation Clustering
To move from pairwise decisions to a global partition of the dataset, we define a Bayesian network that treats clustering as a partition problem. The objective is to find the most probable partition of the data given the pairwise probabilities. Unlike traditional clustering methods like k-means, correlation clustering does not require a pre-defined number of clusters. It treats the problem as a graph where nodes are audio clips and edges represent "same" or "different" predictions. We employ local search heuristics to find high-quality partitions in this NP-hard problem space.
Experimental Setup
The model was evaluated using subsets of the Xeno-Canto database and the ESC-50 environmental noise dataset:
- Train-68: 68 species used for training the Siamese network.
- Test-68: Held-out data from the same 68 species to evaluate performance on "seen" classes.
- Test-0,87: 87 species with minimal data that were entirely absent from the training set, testing "unseen" generalization.
- Test-N: 39 classes of environmental noise (e.g., wind, rain) to assess noise separation capabilities.
Results and Discussion
The performance of various model configurations was measured using metrics such as Variation of Information (VI) and Rand Index (RI). Evaluation shows that applying correlation clustering improves the accuracy of raw pairwise decisions by enforcing transitivity.
Species Agnosticism and Generalization
The Siamese network successfully learned generalizable acoustic features. On the Test-0,87 dataset, the model performed significantly better than random, demonstrating that it understands the concept of species similarity even for birds it has never encountered.
The "Over-splitting" Phenomenon
A notable observation was that the model often produced more clusters than the number of ground-truth species. Analysis suggests this is not necessarily an error but a reflection of intra-species vocal diversity (e.g., distinct songs vs. alarm calls). The model effectively clusters these different vocalization types separately.
Noise Filtering and Decision Refinement
The system proved highly effective at isolating environmental noise, treating it as an outlier that does not cluster with bird sounds. Furthermore, the global correlation clustering process improved the accuracy of raw pairwise decisions by enforcing transitivity—correcting contradictory predictions (e.g., if A matches B and B matches C, the model ensures A matches C).
Conclusion
Correlation clustering offers a robust framework for open-world bird sound discovery. While computationally intensive, the approach provides a pathway for analyzing large-scale bioacoustic data without exhaustive manual labeling. Future work will focus on scaling these heuristics for massive datasets and collaborating with ornithologists to validate the biological significance of the discovered sub-clusters.
Summary updated