PaperlayerView exact revision on arXiv

Source-linked AI summary

Perch 2.0: The Bittern Lesson for Bioacoustics

Bart van Merriënboer, Vincent Dumoulin, Jenny Hamer, Lauren Harrell, Andrea Burns, Tom Denton

arXiv:2508.04665v2cs.LGcs.SDeess.AS

TL;DR

Bioacoustics supports conservation and biodiversity monitoring, and the paper seeks models that can adapt across domains and tasks. Perch 2.0 expands training beyond avian data with additional objectives, achieving state-of-the-art results on two benchmarks. The authors conclude that high-quality transfer does not require very large models, while noting that real-world-representative benchmarks remain an open need.

  • Problem

    Bioacoustics supports conservation and biodiversity monitoring, motivating models that adapt across domains and tasks.

    Source: §1 · p. 2

  • Method

    Perch 2.0 adds non-avian training data and uses prototype-based self-distillation and source prediction alongside species classification.

    Source: §1 · p. 2

  • Results

    Perch 2.0 achieves state-of-the-art results on BirdSET and the BEnchmark of ANimal Sounds (BEANS).

    Source: §1 · p. 1

  • Takeaways & Limitations

    The authors conclude that high-quality transfer learning does not require very large models and that fixed embeddings can be reused across large datasets.

    Source: §4.2 · p. 9

    The authors say more work is needed on benchmarks that reflect real-world use of bioacoustics models.

    Source: §4.3 · pp. 9–10

Abstract

from arXiv · show

Perch is a performant pre-trained model for bioacoustics. It was trained in supervised fashion, providing both off-the-shelf classification scores for thousands of vocalizing species as well as strong embeddings for transfer learning. In this new release, Perch 2.0, we expand from training exclusively on avian species to a large multi-taxa dataset. The model is trained with self-distillation using a prototype-learning classifier as well as a new source-prediction training criterion. Perch 2.0 obtains state-of-the-art performance on the BirdSet and BEANS benchmarks. It also outperforms specialized marine models on marine transfer learning tasks, despite having almost no marine training data. We present hypotheses as to why fine-grained species classification is a particularly robust pre-training task for bioacoustics.

1. Introduction

Perch 2.0 expands Perch beyond avian training and combines supervised species classification with new training objectives. The introduction positions fine-grained classification as a useful pre-training task and emphasizes strong benchmark performance and practical transfer.

1. Introduction

  • The introduction notes that avian-trained models including BirdNET and Perch 1.0 have transferred to individual identification and non-avian bioacoustics tasks.The cited examples include transfer to individual identification and non-avian tasks.Source: §1 · p. 1
  • Perch 2.0 achieves state-of-the-art results on BirdSET and the BEnchmark of ANimal Sounds (BEANS).Source: §1 · p. 1
  • The authors report that their HuBERT and SimCLR experiments, like other self-supervised approaches, did not consistently outperform supervised models.The paper presents hypotheses for why fine-grained avian species classification may be useful for pre-training.Source: §1 · p. 1
  • The paper’s “bittern lesson” is that simple supervised models are difficult to beat.The phrase echoes Sutton’s argument that AI advances come from simple, general-purpose methods.Source: §1 · p. 1
  • The model adds non-avian training data and new augmentations and objectives, including prototype-based self-distillation and source prediction.A multi-source generalization of mixup is introduced; source prediction asks the model to identify the source recording of an audio window.Source: §1 · p. 2
  • Perch aims to support transfer across domains and tasks through linear probing, which requires fewer computational resources, less expertise, and less labeled data than full fine-tuning.The model is based on EfficientNet-B3, described as small by modern machine-learning standards.Source: §1 · p. 2

2. Methodology

Perch 2.0 combines multi-taxa audio data with a compact embedding model trained using species classification, prototype-based self-distillation, and source prediction. Its evaluation covers diverse validation tasks and cross-taxa benchmarks without fine-tuning the embedding model.

2.1. Training data

  • Training combines four labeled audio datasets into 14,795 classes; Xeno-Canto and Tierstimmenarchiv labels are mapped to iNaturalist classes, and bat recordings are removed.The classes include 14,597 species labels and 198 general sound-event classes.Source: §2.1 · p. 2
  • Because recordings have weak labels and variable lengths, training samples fixed 5-second windows using random selection or energy peaks, alongside multi-component mixup with multi-hot targets.The multi-hot targets reflect that all vocalizations in a mixed window should be recognized with high confidence regardless of loudness.Source: §2.1 · p. 3

2.2. Model Architecture

  • The architecture uses an EfficientNet-B3 embedding model and linear, prototype-learning, and source-prediction heads.The embedding network has 12 million parameters; the source-prediction head uses a rank-512 projection.Source: §2.2 · pp. 3–4

2.3. Training objectives

  • Training combines species-classification losses with prototype-to-linear self-distillation and a source-prediction objective, in two phases.The prototype classifier supplies soft targets to the linear classifier; source prediction assigns each recording its own class, inferred from a short window.Source: §2.3 · pp. 4–5

2.4. Hyperparameter selection

  • Vizier selects hyperparameters through separate two-stage sweeps before and during self-distillation, using validation-task performance to choose the model.The first-phase sweep trains 100 models per stage; the second adds the self-distillation loss weight.Source: §2.4 · p. 5

2.5.1. Model-selection tasks

  • Validation tests pretrained classification, one-shot retrieval, and linear transfer across domain-shifted datasets, with the embedding model frozen.The transfer tasks include marine datasets; model selection aggregates performance across task types and datasets using geometric means.Source: §2.5 · p. 6, §2.5.1 · p. 6

2.5.2. Evaluation tasks

  • For evaluation, BirdSet uses the prototype classifier directly, while BEANS trains linear and prototypical probes on embeddings without embedding-model fine-tuning.BEANS includes twelve cross-taxa test tasks; the paper also compares Perch 2.0 with specialized marine models.Source: §2.5.2 · p. 7

3. Results

Perch 2.0 achieves state-of-the-art performance on BirdSet and BEANS without fine-tuning its embedding model. Prototypical probing appears to help on BEANS detection, while random-window training performs on par with energy-peak selection.

3. Results

  • Table 3 reports BirdSet mean cmAP, AUROC, and top-1 accuracy, plus BEANS mean classification accuracy and detection macro-averaged average precision.The full benchmark-subtask breakdown is reported in Appendix C.Source: §3 · p. 7, Table 3 · p. 8
  • In the cited benchmark findings, Perch 2.0 achieves state-of-the-art performance on the BEANS tasks, with this performance obtained without any fine-tuning of the embedding model.Source: §3 · p. 8
  • Prototypical probing seems to improve Perch 2.0’s performance on BEANS detection tasks.Its linear-probing performance on BEANS compares favorably with existing models.Source: §3 · p. 8
  • Random-window training performs on par with energy-peak selection, unlike the authors’ prior Perch 1.0 experience and other reported results.The authors hypothesize that self-distillation helps address label noise, potentially reducing the need for window selection methods.Source: §3 · p. 8

4. Discussion

The discussion argues that supervised bioacoustics pre-training remains strong with abundant labeled data and fine-grained species labels, while reusable embeddings support resource-efficient transfer. It also identifies realistic benchmarks and semi-supervised learning as areas for further work.

4.1. Why supervision?

  • Bioacoustics comparisons find supervised models still outperform self-supervised models, while self-supervised performance depends strongly on domain-specific objectives and augmentations.The authors note that finding suitable bioacoustics augmentations may remain an open issue.Source: §4.1 · p. 9
  • Hundreds of thousands of labeled examples can make supervised models increasingly difficult to outperform, while fine-grained labels improve transfer-learning performance.Perch 2.0 uses over 1.5 million labeled recordings, and reducing label granularity worsens embedding transfer.Source: §4.1 · p. 9

4.2. Conclusion

  • High-quality transfer does not require very large models, and reusable fixed embeddings can reduce resources for large-dataset transfer and one-shot retrieval.The conclusion also describes the model as simple and its embeddings as linearly separable.Source: §4.2 · p. 9

4.3. Future work

  • More work is needed on benchmarks that reflect real-world bioacoustics use, despite validation procedures intended to narrow the evaluation-to-deployment gap.The paper describes deployment as particularly complex.Source: §4.3 · pp. 9–10
  • Source prediction offers an avenue for semi-supervised learning to fill labeled-data gaps, including for taxa with relatively few labeled examples such as mammals.The authors also raise metadata-based classification tasks as a possible direction.Source: §4.3 · p. 10

A.1. Hyperparameters

The appendix reports the selected hyperparameters separately for the first training phase without self-distillation and the second phase.

A.1. Hyperparameters

  • Phase 1 used dropout rate 0.49 and source-prediction loss weight 0.11.Source: §A.1 · p. 14
  • Phase 1 used mixup parameters α=91.3, β=100, ω=1, n=2, and learning rate 6.41 · 10−4.Source: §A.1 · p. 14
  • Phase 2 used self-distillation loss weight 4.22, dropout rate 0, source-prediction loss weight 0, no mixup, and learning rate 3.20 · 10−6.Source: §A.1 · p. 14

A.2. Frontend

The frontend converts five seconds of audio into mel-scaled log-spectrograms using specified framing, windowing, and scaling choices.

A.2. Frontend

  • The frontend takes 5 s of audio at 32 kHz and outputs mel-scaled log-spectrograms.Source: §A.2 · p. 14
  • It uses a 10 ms hop, 20 ms window, 1,024-sample FFT, uncentered frames, and a Hann window.Source: §A.2 · p. 14
  • The frontend computes a magnitude spectrogram, applies HTK mel scaling, then logs values with a 10−5 floor and multiplies by 0.1.Source: §A.2 · p. 14

B. Peak-finding

The peak-finding appendix describes spectrogram construction, threshold-based denoising, peak selection, and fallback handling for short or peakless recordings.

B. Peak-finding

  • The procedure builds a mel-spectrogram with an 80 ms window and 10 ms hop, log-scales magnitudes with a 0.01 floor, and scales them by 0.1.Source: B. Peak-finding · p. 14
  • For each frequency bin, it removes values above the mean plus 1.5 standard deviations, then selects signal above the remaining mean plus 0.75 standard deviations.The selected signal is shifted by the second mean and the rest is discarded.Source: B. Peak-finding · pp. 14–15
  • The method keeps up to five peaks after wavelet-based detection and a 600 ms energy check; recordings under 6 s are zero-padded, and peakless recordings use their first 6 s.Source: B. Peak-finding · p. 15

C.1. Label Granularity

A Xeno-Canto-only model was used to examine how label granularity affects transfer learning, alongside the effect of accumulating additional supervisory data.

C.1. Label Granularity

  • The experiment was designed to demonstrate the importance of label granularity for transfer learning.Source: §C.1 · p. 15
  • The experiment used a Perch 1.0-like architecture without a prototype linear classifier, trained on increased Xeno-Canto data with slightly fewer classes.Some species were removed from Xeno-Canto because of poaching concerns.Source: §C.1 · p. 15
  • Transfer learning performance degrades steadily as the Xeno-Canto training labels become more granular.Source: §C.1 · p. 15
  • The new model trained only on Xeno-Canto outperforms Perch 1.0, demonstrating the impact of accumulated supervisory data.Source: §C.1 · p. 15

C.2. Few shot learning for underwater audio results

Perch 2.0 was evaluated on underwater-audio transfer tasks, where results include strong few-shot performance on marine datasets and a recommendation for cetacean classification.

C.2. Few shot learning for underwater audio results

  • The evaluation covered NOAA PIPAN, ReefSet, and DCLDE 2026; DCLDE was divided into species, ecotype, and known-species label sets.The ecotype set distinguishes five orca ecotypes; the species set also includes abiotic and undetermined biological sounds.Source: §C.2 · p. 15
  • 0.954 AUC-ROC with few-shot embeddings versus 0.612 with direct Google Multispecies logits on DCLDE demonstrates the value of adapting embeddings to specific classes.Source: §C.2 · pp. 17–18, Table 8 · p. 18
  • 0.981 ROC-AUC for Perch 2.0 on ReefSet is close to SurfPerch’s 0.986, although SurfPerch was trained directly on ReefSet.Source: §C.2 · p. 17, Table 8 · p. 18
  • The whale-specific model in the linear probe is recommended for generating embeddings for cetacean-classification tasks.Source: §C.2 · p. 17
  • A separate Xeno-Canto experiment evaluates transfer learning as training labels are converted to different levels of granularity.Source: §C.2 · p. 18, Table 7 · p. 18
  • The marine-transfer table compares ROC-AUC for few-shot learning with 16 training samples per class across Perch and pretrained marine models.An asterisk marks a dataset included in a model’s training data.Source: §C.2 · p. 18, Table 8 · p. 18
Loading 2508.04665v2…