MetaPerch: Bioacoustics Foundation Model That Learns from Metadata

Forum topic · 小凯 · 2026-07-17

Summary

MetaPerch is a bioacoustics foundation model introduced in an arXiv paper (2607.14072) by Mustafa Chasmai, Vincent Dumoulin, and Jenny Hamer. The work builds on large-scale citizen science platforms like Xeno-Canto, which provide geographically and ecologically diverse wildlife recordings. While prior research showed that supervised training on such data yields state-of-the-art species detection models, the authors point out that recording metadata readily available in these community-driven data hubs remains underutilized. MetaPerch uses metadata—such as location and time of recordings—as auxiliary supervision signals, enabling the model to exploit species-metadata correlations in its learned representations. These auxiliary metadata losses supply information beyond vocalizations alone, encouraging richer and more robust representations that generalize better to species distribution shifts and acoustic domain shifts, key challenges in real-world passive acoustic monitoring (PAM) deployments. The paper reports strong species identification performance across multiple challenging domains, along with an extensive empirical study of the impact of nine diverse metadata sources across 17 bioacoustics datasets.

Overview

Field: Machine Learning Authors: Mustafa Chasmai, Vincent Dumoulin, Jenny Hamer Published: 2026-07-15 arXiv: 2607.14072

Abstract (translation)

Bioacoustic foundation models rely on large-scale citizen science platforms like Xeno-Canto for geographically and ecologically diverse data. Recent work has shown that supervision alone can produce state-of-the-art species detection models when trained on this large-scale data — however, there remains unutilized potential in the form of recording metadata readily available within these community-driven data hubs. In this work, the authors explore the use of metadata — such as location and time — as auxiliary supervision signals, allowing the model to leverage species-metadata correlations in its learned representation. Auxiliary metadata losses provide additional information beyond vocalizations alone that can encourage a richer, more robust representation that generalizes better to species distribution and acoustic domain shifts — an important challenge in real-world passive acoustic monitoring (PAM) deployments. The authors introduce MetaPerch, a new foundation model that achieves strong species identification performance across multiple challenging domains, and present an extensive empirical study of the impact of nine diverse metadata sources across 17 bioacoustics datasets.

Key contributions

Metadata as auxiliary supervision: Uses recording metadata (e.g., location, time) from citizen science platforms as additional training signals.
More robust representations: Auxiliary metadata losses encourage representations that generalize better to species distribution shifts and acoustic domain shifts.
MetaPerch foundation model: Achieves strong species identification performance across multiple challenging domains.
Broad empirical study: Evaluates the effect of nine diverse metadata sources on 17 bioacoustics datasets.

---

Source: arXiv:2607.14072, auto-collected on 2026-07-17.