rfcx/frugalai

Dataset Card for RFCx Chainsaw Audio

Dataset Summary

A large set of short audio clips of chainsaws at varying distances. The data comes from Guardian devices deployed by Rainforest Connection to detect illegal logging. The majority of the recordings are from South America or South East Asia where Rainforest Connection has a large number of projects.

Supported Tasks and Leaderboards

This dataset is made for binary audio classification.

Dataset Structure

The dataset only contains audio data and labels. Each audio clip is 3 seconds long. Each sample is labelled either chainsaw (value 0 - positively identifying a chainsaw) or environment (value 1 - not containing a chainsaw).

Data Instances

{
    'audio': {
        'path': 'aoos_2021_02a16dd4-c788-4bbb-bc3d-e2f8322fe4b2_0-3.wav',
        'array': array([4.84344482e-01, 4.54193115e-01, 2.53906250e-02, ..., 2.44140625e-04, 3.05175781e-05, 9.15527344e-04]),
        'sampling_rate': 12000
    },
    'label': 0
}

Data Fields

• audio.path the filename of the audio file
• audio.array the raw waveform of the audio
• audio.sampling_rate the sample rate of the audio
• label chainsaw 0 or environment 1

Data Splits

The data is split into train and test sets.

Dataset Creation

Curation Rationale

The chainsaw dataset was compiled to train a model that could run on devices in the forest and detect illegal logging in real-time. The devices send a message to rangers on the ground to make an intervention.

Source Data

The data is from Guardian devices deployed between 2015 and 2022. The audio was typically recorded at 12kHz and saved in opus format (lossy compression) before upload to the cloud over EDGE or 3G networks. For more information, see Guardian device software.

Annotations

Annotation process

Where Guardians are deployed, rangers and field teams responsible for monitoring alerts can confirm and reject detected chainsaws via the mobile Guardian app or via the Guardian dashboard web app.

Who are the annotators?

The field team and partners of Rainforest Connection.

Personal and Sensitive Information

The source data typically originates from deep in the forest, from a device high up in the forest canopy, hence human voices are extremely rare. Where human voices have been found in existing datasets, those samples have been removed.

Getting Started

Example, in a colab, install the dependencies:

!pip install librosa soundfile datasets

Load the dataset and print one example:

from datasets import load_dataset
dataset = load_dataset("rfcx/frugalai", streaming=True)
print(next(iter(dataset['train'])))

See HuggingFace's audio documentation for more examples.

Additional Information

Licensing Information

The dataset is provided under the CC BY-NC 4.0 license.