Natural Sound Geolocation Benchmark

This repo provides benchmark code for audio geolocation on iNatsounds and Xeno-Canto Dawn Chorus (XCDC) datasets, and also includes the official release of the XCDC dataset.

Figure 1: iNatSounds and XCDC predictions from around the world.

iNatSounds

Please see visipedia/inat_sounds for downloading instructions and documentation for iNatSounds. The dataset json files (for train, val and test) include latitude and longitude for each audio, which serve as labels for geolcoation.

The XCDC Dataset

We present the XenoCanto Dawn Chorus (XCDC) Dataset, a new benchmark for audio geolocation using wildlife sounds. Dawn chorus is a symphony of bird songs involving complex sounds. These communal vocal displays occur at dawn and are most noticeable during the breeding season. A rich diversity of species can be heard in a relatively short amount of time, possibly providing the cues needed to narrow down the location of a recording. We construct a new benchmark to test this hypothesis. From XenoCanto, we select audio that:

• was recorded during spring-time dawn chorus.
• is at least 3 minutes long.
• contains at least 10 distinct species.

This yields 576 recordings with corresponding geographic locations and species annotations.

We release all preprocessed audio .wav files and their annotations here. The annotation also contains urls to the original observation in XenoCanto, which can be used to access raw recordings.

Annotation Format

The xcdc_recordings.csv file contains metadata and labels for the dataset. Each row corresponds to one data sample. Below is a description of each column.

Column Name	Description
audio_id	Unique ID for each audio recording
url	Link to the original audio on Xeno-Canto
latitude	Latitude coordinate where the audio was recorded
longitude	Longitude coordinate where the audio was recorded
length	Duration of the recording (MM:SS or HH:MM:SS)
date	Recording date
time	Recording time
recordist	Name of the person who recorded the audio
main_scientific	Scientific name of the primary specie present in the recording
main_common	Common name of the primary specie present in the recording
all_scientific	List of scientific names of all annotated species in the recording
all_common	List of common names of all annotated species in the recording

Note: A Croissant metadata file xcdc_croissant.json is also provided. To integrate it into your machine learning workflow, please refer to Hugging Face Croissant Guide.

Benchmarking Geolocation

To benchmark your method on these two datasets, please create a prediction csv file with the following two fields. See the ground-truth files for XCDC and iNatSounds for examples.

Column Name	Description
audio_id	Unique ID for each audio recording. This is also the filename of the audio recording (basename, without the extension)
latitude	Predicted Latitude coordinate
longitude	Predicted Longitude coordinate

With this csv, run

python3 evaluate_geo.py --pred_cvs <> --dataset <>

where the dataset argument can be xcdc or inat. This should give you a list of different metrics as reported in the paper.

Citation

If you use the dataset and benchmark in your work, please consider citing the following papers:

@article{chasmai2024inaturalist,
    title={The iNaturalist Sounds Dataset},
    author={Chasmai, Mustafa and Shepard, Alexander and Maji, Subhransu and Van Horn, Grant},
    journal={Advances in Neural Information Processing Systems},
    volume={37},
    pages={132524--132544},
    year={2024}
}