Darksound
Authors/Creators
Description
Darksound is an open-source and code-based dataset for the evaluation of unsupervised meta-learning algorithms in the context of ecoacoustics. This dataset is composed of regions of interest (ROIs) of nocturnal and crepuscular bird species living in tropical environments that were automatically segmented using the Python package Bambird (Michaud et al., 2023). The dataset is split into two sets, with a training set and a test set.
Technical info (English)
Data acquisition
Data acquisition for building the dataset was made using the recordings freely available from the Xeno-Canto database. All the recordings were in the mp3 format with varying sampling rates ranging from 8 kHz to 96 kHz. Recordings were resampled to a sampling rate of 44.1 kHz, with the aliasing problem of the recordings resolved using a high-quality Fast Fourier Transform-based bandlimited interpolation.
Training set
The training set was built by sending query parameters through the API of Xeno-Canto which returned a JSON object containing recording metadata. The recordings were downloaded according to four query parameters: (i) audio quality, corresponding to the highest audio quality (i.e. A quality level), (ii) duration, which corresponded to recordings lasting from 20 to 60 seconds, (iii) maximum number of recordings allowed per species, which was set to 100, and (iv) geographic coordinates of the recordings that surrounded the Equator line in America. Geographical coordinates were defined according to the latitude of the Tropics, with the Tropic of Cancer in the Northern Hemisphere at 23° 26 '10.6”N and the Tropic of Capricorn in the Southern Hemisphere at 23° 26' 10.6”S. This included an initial training set of 29,587 audio recordings corresponding to 2,976 species of tropical bird species.
Test set
The test set was built in the same manner as the training set, except that it used (i) both high and low quality for the Xeno-Canto recordings (i.e. from A (highest quality) to E (lowest quality) quality levels), and (ii) no restrictions for the length of the recording in seconds. The query included recordings of 30 nocturnal and crepuscular tropical bird species that were not present in the training set. Tropical bird species were tinamous (Tinamidae), nocturnal raptors (Falconidae, Strigidae and Nyctibiidae), and nightjars (Caprimulgidae) which are respectively considered as “umbrella” and “sentinel” species of the Amazonian forest. This included an initial test set of 2,106 audio recordings.
(NB: Because the test set has been manually labeled, audio recordings are here already segmented. Segmentation process was accomplished using the Bambird package with the default parameters, except for the frequency bandpass filter which was applied between 250 and 5,000 Hz to each recording, so that bird sounds could be kept and insect sounds excluded. All ROIs were faded in and out to avoid aliasing effects due to edge effects, and zero-padded to a maximum duration of 3 seconds to obtain ROIs of the same size. This resulted in a test set containing 30 nocturnal and crepuscular tropical bird species for a total number of 8,163 ROIs manually labeled.)