Dataset card for NatureLM-audio-training

Overview

NatureLM-audio-training is a large and diverse audio-language dataset designed for training bioacoustic models that can generate a natural language answer to a natural language query on a reference bioacoustic audio recording. For example, for an in-the-wild audio recording of a bird species, a relevant query might be "What is the common name for the focal species in the audio?" to which an audio-language model trained on this dataset may respond with "Common yellowthroat".

It consists of over 26 million audio-text pairs derived from diverse sources including animal vocalizations, insects, human speech, music and environmental sounds:

Task	Dataset	# Hours	# Samples
CAP	WavCaps (Mei et al., 2023)	7,568	402k
CAP	AudioCaps (Kim et al., 2019)	145	52k
CLS	NSynth (Engel et al., 2017)	442	300k
CLS	LibriSpeechTTS (Zen et al., 2019), VCTK (Yamagishi et al. 2019)	689	337k
CAP	Clotho (Drossos et al. 2020)	25	4k
CLS, DET, CAP	Xeno-canto (Vellinga & Planque, 2015)	10,416	607k
CLS, DET, CAP	iNaturalist (iNaturalist)	1,539	320k
CLS, DET, CAP	Watkins (Sayigh et al., 2016)	27	15k
CLS, DET	Animal Sound Archive (Museum für Naturkunde Berlin)	78	16k
DET	Sapsucker Woods (Kahl et al., 2022a)	285	342k
CLS, DET	Barkley Canyon (Kanes, 2021)	876	309k
CLS	UrbanSound (Salamon & Jacoby, 2014)	10	2k

CLS: classification, DET: detection, CAP: captioning, Samples = number of audio files

Introduced in the paper NatureLM-audio: An Audio-Language Foundation Model for Bioacoustics, this dataset aggregates data from sources such as Xeno-canto, iNaturalist, Animal Sound Archive amongst others. The accompanying model trained on this dataset can be found here EarthSpeciesProject/NatureLM-audio.

• Developed by: David Robinson, Marius Miron, Masato Hagiwara, Milad Alizadeh, Gagan Narula, Sara Keen, Benno Weck, Matthieu Geist, Olivier Pietquin (Earth Species Project)
• Funded by: More info at https://www.earthspecies.org/about-us#support
• Shared by: Earth Species Project
• Language(s) (NLP): English

Coverage of taxonomic groups

The NatureLM-audio-training dataset, though diverse, still leans toward bird species.

Usage

from datasets import load_dataset

dataset = load_dataset("EarthSpeciesProject/NatureLM-audio-training", split="train")
print(dataset)

Example data

import numpy as np

# Inspect the first example in the dataset
x = ds[0]
audio = x["audio"]["array"]
print(audio.shape)  
# (503808,)

print(x["instruction"])  
# '<Audio><AudioHere></Audio> What is the taxonomic name of the focal species in the audio?'

print(x["output"])
# 'Chordata Aves Passeriformes Passerellidae Atlapetes fuscoolivaceus'

print(x["task"])
# 'taxonomic-classification'

import json
metadata = json.loads(x["metadata"])
print(metadata)
# {'recordist': 'Peter Boesman',
#  'url': 'https://xeno-canto.org/sounds/uploaded/OOECIWCSWV/XC693334-LS_58842%20Dusky-headed%20brushfinch%20song%20B.mp3',
#  'source': 'Xeno-canto',
#  'duration': 31.488,
#  'class': 'Aves',
#  'family': 'Passerellidae',
#  'genus': 'Atlapetes',
#  'species': 'Atlapetes fuscoolivaceus',
#  'phylum': 'Chordata',
#  'order': 'Passeriformes',
#  'subspecies': '',
#  'data_category': 'animal',
#  'text': None,
#  'sample_rate': 16000}

Example prompts

Prompt: What is the common name for the focal species in the audio?
Answer: Humpback Whale

Prompt: Which of these, if any, are present in the audio recording? Single pulse gibbon call, Multiple pulse gibbon call, Gibbon duet, None.
Answer: Gibbon duet

Prompt: What is the common name for the focal species in the audio?
Answer: Spectacled Tetraka

Prompt: What is the life stage of the focal species in the audio?
Answer: Juvenile

Prompt: What type of vocalization is heard from the focal species in the audio?
Answer with either 'call' or 'song'.

Prompt: Caption the audio, using the common name for any animal species.

Dataset Composition

NatureLM-audio-training combines data from several well-known sources. There are a total of 26,440,512 samples (examples). The data are organized in shards of 2500 samples each, each containing a subset of the total dataset. We additionally provide an annotations.jsonl file that contains taxonomic information for each sample, including family, genus, species, common name, and other relevant metadata. This file can be used to perform queries on the dataset to create data mixes as desired by the user. The id column of the annotations.jsonl file matches the id field in the dataset, and the shard-id of the corresponding sample is in the shard_id column.

Tasks and Applications

Several different types of bioacoustically relevant tasks are included in the dataset. The tasks are designed to be flexible and can be used for a variety of applications, including (but not limited to):

• taxonomic-classification
• species-sci-detection-hard
• genus-detection
• call-type
• caption-scientific-rich
• open-ended question
• call-type-with-common-name
• lifestage-with-common-name ...

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 (Audio): The audio data in float32 format.
• id (str): Sample uuid.
• metadata (str): Each sample can have some extra data such as the url of the original audio, the recordist, duration, sample_rate and taxonomic information such as family, genus, species, common name, etc. The metadata is a JSON string.
• 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. 'taxonomic-classification', 'caption-common', 'lifestage', 'speech-nspeakers'
• license (str): The license of the dataset. For example, 'CC-BY-NC' or 'CC0'.

Licensing

Due to its composite nature, NatureLM-audio-training 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 NatureLM-audio-training, 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}
}

Contact

For questions, comments, or contributions, please contact:

• D. Robinson (david at earthspecies dot org)
• M. Hagiwara (masato 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)