Model Card for esp-aves2-sl-beats-all

Model Details

Model Description

esp-aves2-sl-beats-all is an audio representation learning model (bioacoustic encoder) designed to produce transferable embeddings for downstream bioacoustic tasks including species classification and detection, individual identification, and vocal repertoire discovery, as described in What Matters for Bioacoustic Encoding.

• Developed by: Marius Miron, David Robinson, Milad Alizadeh, Ellen Gilsenan-McMahon, Gagan Narula, Emmanuel Chemla, Maddie Cusimano, Felix Effenberger, Masato Hagiwara, Benjamin Hoffman, Sara Keen, Diane Kim, Jane K. Lawton, Jen-Yu Liu, Aza Raskin, Olivier Pietquin, Matthieu Geist
• Funded by: More info at https://www.earthspecies.org/about-us#support
• Shared by: Earth Species Project
• Model type: Audio representation learning model (Transformer; BEATs backbone)
• License: CC-BY-NC-SA
• Finetuned from model: BEATs pretrained on AudioSet (see Parent Models)

Model Sources

• Repository: https://github.com/earthspecies/avex
• Paper: What Matters for Bioacoustic Encoding
• Hugging Face Model: ESP-AVES2 Collection
• Configuration: train_config.yaml

Parent Models

This model is based on or fine-tuned from the following parent models:

1. BEATs (pretrained on AudioSet)
   • Source: https://github.com/microsoft/unilm/tree/master/beats
   • Description: Self-supervised transformer audio encoder used as the base SSL checkpoint.
   • License: See upstream repository

Uses

Direct Use

esp-aves2-sl-beats-all can be used directly for bioacoustic tasks such as species classification and detection, repertoire and individual classification, retrieval and clustering of audio.

Downstream Use

The model can be used for linear probing, retrieval, and clustering of audio; it can also be fine-tuned for task- and domain-specific bioacoustic applications (taxon-, habitat-, or device-specific).

Out-of-Scope Use

The model is not designed as a generative model, and it does not produce text outputs. Using it as a stand-alone classifier without training a probe or finetuning is out of scope.

Bias, Risks, and Limitations

• Bias: The training data relies heavily on citizen-science recordings and may over-represent certain taxa and regions (e.g., Northern Hemisphere); this can impact generalization.
• Risks: Predictions and embeddings can be misused for harmful wildlife exploitation (e.g., locating endangered species) if deployed without safeguards.
• Limitations: The paper evaluates and trains models at 16 kHz for fairness; some taxa may require higher bandwidth. Performance can degrade under large distribution shifts (habitat, device, background noise).

Recommendations

Use esp-aves2-sl-beats-all as an encoder (feature extractor) and validate performance on your target domain. For sensitive deployments, apply access controls and follow conservation best practices.

How to Get Started with the Model

Loading this model requires the AVEX (Animal Vocalization Encoder) library avex to be installed.

Installation

pip install avex

Or with uv:

uv add avex

For more details, see https://github.com/earthspecies/avex.

Loading the Model

from avex import load_model

# Model config name depends on how the checkpoint is packaged for this repo.
# If/when an official config is provided, replace the string below accordingly.
model = load_model("esp_aves2_sl_beats_all", device="cuda")

Using the Model

# Case 1: embedding extraction (features only)
backbone = load_model("esp_aves2_sl_beats_all", device="cuda", return_features_only=True)

with torch.no_grad():
    embeddings = backbone(audio_tensor)
    # Shape: (batch, time_steps, 768) for BEATs

# Pool to get fixed-size embedding
embedding = embeddings.mean(dim=1)  # Shape: (batch, 768)

# Case 2: supervised predictions (logits over label IDs; see label_map.json)
model = load_model("esp_aves2_sl_beats_all", device="cuda")

with torch.no_grad():
    logits = model(audio_tensor)
    predicted_class = logits.argmax(dim=-1).item()

Transfer Learning with Probes

from avex.models.probes import build_probe_from_config
from avex.configs import ProbeConfig

# Load backbone for feature extraction
base = load_model("esp_aves2_sl_beats_all", return_features_only=True, device="cuda")

# Define a probe head for your task
probe_config = ProbeConfig(
    probe_type="linear",
    target_layers=["last_layer"],
    aggregation="mean",
    freeze_backbone=True,
    online_training=True,
)

probe = build_probe_from_config(
    probe_config=probe_config,
    base_model=base,
    num_classes=10,  # Your number of classes
    device="cuda",
)

Class Label Mapping

The class label mapping for this supervised learning model can be found at label_map.json in the Hugging Face repository.

Training Details

Training Data

From the paper, the model uses a two-stage recipe: a BEATs SSL backbone pretrained on AudioSet, followed by supervised post-training on an All mix (Bioacoustics mix + AudioSet).

Training Data Sources

Dataset	Description	Source	License	Size
AudioSet	general audio	Link	See dataset terms	5700 hours
Xeno-canto	birds	Link	CC (varies)	10416 hours
iNaturalist	diverse taxa	Link	CC (varies)	1539 hours
Watkins	marine mammals	Link	licensing agreement (paper)	27 hours
Animal Sound Archive	diverse taxa	Link	See archive terms	78 hours

Training Procedure

As described in the paper:

• Stage 1 (SSL): BEATs pretrained on AudioSet.
• Stage 2 (SL): supervised post-training on All (Bio + AudioSet).
• Augmentations: random additive noise with probability 0.5 at SNR sampled uniformly from ([-10, 20]) dB; mixup-style linear mixing of random pairs in-batch with probability 0.5 and union of labels.

Training Hyperparameters

Training hyperparameters are specified in train_config.yaml.

Evaluation

Testing Data, Factors & Metrics

Testing Data

The paper evaluates on a benchmark spanning:

• BEANS (classification and detection): https://github.com/earthspecies/beans
• BirdSet (detection; Dedicated Train setup): https://huggingface.co/datasets/DBD-research-group/BirdSet
• Individual ID (classification): Pipit, Chiffchaff, Little Owl, Macaques
• Vocal Repertoire (retrieval + clustering): Zebra Finch, Giant Otters, Bengalese Finch, Killer Whale

Metrics

The paper reports:

• Linear probing: accuracy (single-label) and mean average precision (multi-label/detection)
• Retrieval: ROC AUC
• Clustering: normalized mutual information (NMI) for single-label datasets

Results

Aggregate results for linear probing (frozen base model) with esp-aves2-sl-beats-all (from the provided LaTeX table):

Benchmark	Task	Metric	Score
BEANS Classification	Probe	Accuracy	0.832
BEANS Classification	Retrieval	ROC AUC	0.813
BEANS Classification	Clustering	NMI	0.604
BEANS Detection	Probe	mAP	0.408
BEANS Detection	Retrieval	ROC AUC	0.726
BirdSet	Probe	mAP	0.294
BirdSet	Retrieval	ROC AUC	0.732
Individual ID	Probe	Accuracy	0.511
Individual ID	Retrieval	ROC AUC	0.690
Vocal Repertoire	Retrieval	ROC AUC	0.798
Vocal Repertoire	Clustering	NMI	0.529

Environmental Impact

Not specified in the provided excerpt.

Technical Specifications

Model Architecture and Objective

esp-aves2-sl-beats-all uses a BEATs transformer encoder trained with a self-supervised pretraining stage (AudioSet) followed by supervised post-training on All (Bio + AudioSet), to learn general-purpose bioacoustic representations.

Key components:

• Encoder: BEATs transformer
• Feature extraction: time-series embeddings, pooled for probes/retrieval/clustering in the paper
• Output: embeddings (dimension depends on backbone configuration)

Compute Infrastructure

Not specified in the provided excerpt.

Model Configuration

Model configuration is available in train_config.yaml.

Citation

BibTeX:

@inproceedings{miron2025matters,
  title={What Matters for Bioacoustic Encoding},
  author={Miron, Marius and Robinson, David and Alizadeh, Milad and Gilsenan-McMahon, Ellen and Narula, Gagan and Chemla, Emmanuel and Cusimano, Maddie and Effenberger, Felix and Hagiwara, Masato and Hoffman, Benjamin and Keen, Sara and Kim, Diane and Lawton, Jane K. and Liu, Jen-Yu and Raskin, Aza and Pietquin, Olivier and Geist, Matthieu},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026}
}

APA:

Miron, M., Robinson, D., Alizadeh, M., et al. (2025). What Matters for Bioacoustic Encoding. arXiv preprint arXiv:2508.11845.

Glossary

• Bioacoustic encoder: A model that maps audio to embeddings useful for downstream bioacoustic tasks.
• Linear probing: Training a simple linear model on frozen embeddings to assess representation quality.
• NMI: Normalized Mutual Information, a clustering quality metric.

More Information

• Project page: TBA
• Documentation: TBA
• Issue tracker: https://github.com/earthspecies/avex/issues

Model Card Authors

• Earth Species Project

Model Card Contact

Contact: marius@earthspecies.org, david@earthspecies.org, milad@earthspecies.org, gagan@earthspecies.org