Hugging Face's logo

Datasets:
EarthSpeciesProject
/
BEANS-Zero

Tasks:
Audio Classification
Modalities:
Text
Formats:
parquet
Languages:
English
Size:
10K - 100K
ArXiv:
Tags:
biology
bioacoustics
audio-classification
multimodal
zero-shot
Libraries:
Datasets
Dask
License:
Dataset card Data Studio Files
xet
Community
1
BEANS-Zero / README.md
gagannarula's picture
gagannarula
still testing
c30a28b verified
preview code
|
raw
history blame
6.18 kB
metadata
license: other
task_categories:
  - audio-classification
language:
  - en
tags:
  - biology
  - bioacoustics
  - audio-classification
  - multimodal
  - zero-shot
pretty_name: BEANS-Zero
size_categories:
  - 100K<n<1M
configs:
  - config_name: BEANS-Zero
    data_files:
      - split: test
        path: shard*

BEANS-Zero

Version: 0.1.0 Created on: 2025-04-12 Creators:

Overview

BEANS-Zero is a benchmark bioacoustic dataset designed for zero-shot bioacoustic evaluation tasks. Introduced in the paper NATURELM-AUDIO: AN AUDIO-LANGUAGE FOUNDATION MODEL FOR BIOACOUSTICS, this dataset aggregates multiple subsets from both established component datasets and newly curated ones. It is a benchmark dataset only to be used to evaluate audio-text multimodal models that accept a bioacoustic audio input (query audio) and a text prompt (e.g. 'What species is in this audio?') and generate text as output (e.g. "Taeniopygia guttata").

Usage 

import numpy as np
from datasets import load_dataset

ds = load_dataset("EarthSpeciesProject/BEANS-Zero", split="test") # set streaming=True if you want to stream 

# see the contents at a glance
print(ds)

# get audio for a sample
audio = np.array(ds[0]["audio"])
print(audio.shape)

# get the instruction (prompt / query) for that sample
print(ds[0]["instruction_text"])
# the desired output (*only* used for evaluate)
print(ds[0]["output"])

Dataset Composition

BEANS-Zero combines data from several well-known sources. There are total of 113,679 samples (examples). It consists of two main groups:

Already Part of the BEANS Benchmark

  • esc-50: A labeled collection of 2000 environmental audio recordings (5 seconds each) spanning 50 classes. (CC-BY-NC)
  • Watkins: Marine mammal sound recordings.
  • CBI: Cornell Birdcall Identification dataset. (CC-BY-NC-SA)
  • HumBugDB: A large-scale, multi-species dataset of mosquito sounds. (CC-BY)
  • Enabirds: Bird dawn chorus detection data. (CC0)
  • HICEAS: Marine mammal vocalizations from the Hawaiian Islands.
  • RFCX: Recordings for bird and frog vocalizations in soundscape recordings (academic/research & non-commercial use).
  • Hainan Gibbons: Automated detection of Hainan gibbon calls. (CC-BY-NC-SA)

Newly Added Subsets

  • Lifestage: Annotations for classifying animal lifestages (adult, juvenile, nestling).
  • Call-type: Labels distinguishing between call and song in animal sounds.
  • Unseen-species-cmn / sci / tax: Tasks for classifying species common names, scientific names, and taxonomic names not seen during training.
  • Unseen-genus-cmn / sci / tax: Tasks for classifying genus-level names (common, scientific, taxonomic) of unseen species.
  • Unseen-family-cmn / sci / tax: Tasks for classifying family-level names (common, scientific, taxonomic) of unseen species.
  • Captioning: English captions for bioacoustic recordings.
  • ZF-Indiv: A zebra finch dataset with individual and call-type annotations.

Each subset has its own metadata and licensing details. Each sample

Sources

BEANS-Zero was assembled using data from:

  • Xeno-canto
  • iNaturalist
  • Animal Sound Archive
  • Elie et al 2020
  • ESC-50
  • RFCX
  • CBI
  • HumBugDB
  • Enabirds
  • HICEAS
  • Watkins
  • Hainan Gibbons

Tasks and Applications

BEANS-Zero supports various research tasks:

  • Audio Classification: Identify and categorize animal sounds.
  • Audio Detection: Detect specific sound events in recordings.
  • Audio Captioning: Generate natural language descriptions of bioacoustic recordings.

These tasks are particularly useful for exploring zero-shot learning applications in bioacoustics.

Data Fields

The following fields are present in each example:

  • source_dataset (str): One of the source datasets mentioned above
  • audio (Sequence[float]): The audio data in float32 format. The audio is not decoded, and the sample rate is 16000 Hz.
  • id (str): Sample uuid.
  • created_at (str): Sample creation datetime in utc
  • metadata (str): Each sample can have a different duration
  • file_name (str): Sample file_name
  • instruction (str): A prompt (a query) corresponding to the audio for your audio-text model with a placeholder for audio tokens. E.g. ' What is the scientific name for the focal species in the audio?'
  • instruction_text (str): Same as instruction but without the placeholder for audio tokens.
  • output (str): The expected output from the model
  • task (str): The task type e.g. classification / detection / pitch estimation / captioning.
  • dataset_name (str): Names corresponding to the evaluation tasks, e.g. 'esc50' or 'unseen-family-sci'.
  • license (str): The license of the dataset. For example, 'CC-BY-NC' or 'CC0'.

Licensing

Due to its composite nature, BEANS-Zero is subject to multiple licenses. Individual samples have the "license" field indicating the specific license for that sample. The dataset is not intended for commercial use, and users should adhere to the licenses of the individual datasets.

Citation

If you use BEANS-Zero, please cite the following:

@misc{naturelm-audio,
  title={NATURELM-AUDIO: AN AUDIO-LANGUAGE FOUNDATION MODEL FOR BIOACOUSTICS},
  url={https://arxiv.org/pdf/2411.07186},
  note={Preprint},
  year={2024}
}

How to Use

BEANS-Zero is provided with a default configuration. Data files are belong to a sing

Contact

For questions, comments, or contributions, please contact:

  • M. Hagiwara (masato at earthspecies dot org)
  • D. Robinson (david at earthspecies dot org)
  • M. Miron (marius at earthspecies dot org)
  • G. Narula (gagan at earthspecies dot org)
  • M. Alizadeh (milad at earthspecies dot org)