[Paper Review] Detecting bird sound in unknown acoustic background using crowdsourced training data

Timos Papadopoulos, Stephen Roberts+1 more|arXiv (Cornell University)|May 24, 2015

Animal Vocal Communication and Behavior4 citations

TL;DR

This paper proposes a scalable, unsupervised approach to detect bird vocalizations in unknown acoustic backgrounds using crowdsourced audio data from xeno-canto. It trains generative Gaussian Mixture Models (GMMs) on spectrogram features from open-access recordings, achieving effective novelty detection without manual preprocessing, with consistent performance across diverse environments and signal-to-noise ratios, though performance varies by species due to spectral overlap with background sounds.

ABSTRACT

Biodiversity monitoring using audio recordings is achievable at a truly global scale via large-scale deployment of inexpensive, unattended recording stations or by large-scale crowdsourcing using recording and species recognition on mobile devices. The ability, however, to reliably identify vocalising animal species is limited by the fact that acoustic signatures of interest in such recordings are typically embedded in a diverse and complex acoustic background. To avoid the problems associated with modelling such backgrounds, we build generative models of bird sounds and use the concept of novelty detection to screen recordings to detect sections of data which are likely bird vocalisations. We present detection results against various acoustic environments and different signal-to-noise ratios. We discuss the issues related to selecting the cost function and setting detection thresholds in such algorithms. Our methods are designed to be scalable and automatically applicable to arbitrary selections of species depending on the specific geographic region and time period of deployment.

Motivation & Objective

• To develop a scalable method for detecting bird vocalizations in complex, unknown acoustic environments without relying on manually annotated background data.
• To leverage crowdsourced, minimally annotated audio from xeno-canto to train species-specific normality models for bird sounds.
• To evaluate the effectiveness of generative models in distinguishing bird vocalizations from diverse non-bird sound backgrounds, including urban and natural settings.
• To enable automatic, species-specific detection that can be integrated into larger classification pipelines or human-in-the-loop validation systems.
• To assess the impact of feature selection and model complexity on detection performance across different acoustic conditions.

Proposed method

• Training generative models using Gaussian Mixture Models (GMMs) on spectrogram frames extracted from open-access bird recordings obtained from xeno-canto.
• Using a combination of spectral flatness measure (SFM), spectral centroid, and spectral flux as feature vectors to represent bird vocalizations.
• Applying the Minimum Description Length (MDL) criterion to determine the optimal number of GMM components, ensuring model complexity is balanced with data fit.
• Employing a novelty detection framework where low likelihood of a test segment under the learned GMM indicates potential bird vocalization.
• Using a probabilistic thresholding approach based on Extreme Value Theory to set decision boundaries between normal (bird-like) and abnormal (non-bird) segments.
• Testing the method on real environmental recordings from the IEEE AASP Challenge database, including 'Park' and 'Open air market' categories with varying signal-to-noise ratios.

Experimental results

Research questions

• RQ1 Can generative models trained on crowdsourced, minimally annotated bird recordings effectively detect bird vocalizations in unknown acoustic backgrounds?
• RQ2 How does the choice of audio features (e.g., SFM, spectral centroid, spectral flux) affect detection performance across different species and environments?
• RQ3 To what extent does background complexity (e.g., urban vs. natural environments) and signal-to-noise ratio degrade detection performance?
• RQ4 Can the MDL criterion reliably determine the optimal number of GMM components without overfitting, especially when training data are limited?
• RQ5 How can a principled thresholding method be applied to GMM-based novelty detection to balance sensitivity and specificity in real-world bioacoustic monitoring?

Key findings

• The method achieves consistent detection performance across different random data splits and background environments, indicating robustness and reproducibility.
• The inclusion of spectral flatness measure (SFM) as a feature improves detection for certain species, such as *Turdus merula* and *Turdus philomelos*, suggesting spectral texture is informative.
• Performance degrades moderately when transitioning from the 'Park' to the 'Open air market' background, and further when signal-to-noise ratio is reduced by 6dB, though the drop is limited.
• The MDL criterion yields consistent optimal GMM component counts (typically 5–10) across species and feature sets, and increasing components beyond this range does not improve performance.
• Species with low-frequency dominant energy in the mode feature (e.g., 1–2 kHz) show poorer discrimination, as these frequencies overlap with human speech and urban noise.
• The baseline method underperforms for species like *Luscinia luscinia*, *Luscinia megarhynchos*, and *Turdus merula*, likely due to spectral similarity with background sounds, indicating a need for more discriminative features or temporal modeling.