Datasets for "Bioacoustic raw-audio transformer is invariant to input sample rate"

Read full text Summarise with VIRAT
Authors: Marius Faiß
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This repository contains a dataset of audio files containing meerkat and bird vocalizations used for training machine learning classifiers. It was created for the study "Bioacoustic raw-audio transformer is invariant to input sample rate", which has been submitted for publication. The dataset was compiled from publicly available sources, namely the MeerKAT: Meerkat Kalahari Audio Transcripts dataset and sound recordings of the 11 most common bird species on the public database xeno-canto. The dataset contains audio files at varying sample rates, ranging from 8 to 48 kHz. It contains 378323 audio files in total, with label files that store the annotations of sound events for each file. The dataset is 1026 hours in length, with one half stemming from the MeerKAT dataset (11 sound classes) and the other half from xeno-canto (11 species). The audio files are located in the folder "wav", with several subdirectories. The audio files are stored in .mp3 and .wav formats. The folder "lbl" has the same structure and stores the accompanying label files in .h5 format. The "manifests" folder contains the manifest files used for training the animal2vec model. An accompanying csv file lists the contributors of each audio file sourced from xeno-canto. In our study "Bioacoustic raw-audio transformer is invariant to input sample rate", we used an upsampled version of the dataset. Due to storage size restrictions, we only publish the source files at their original sample rates, but provide the scripts that were used to upsample the dataset for repeatability purposes. The code is publicly available at the studies' GitHub repository. Further information on the dataset curation process can be found in the publication, once it is published. The code used for curating the dataset is also available on the GitHub repository. As in the original MeerKAT dataset, by agreement with the Kalahari Research Centre (KRC), we have made these data available in a way that can further machine learning research without compromising the ability of the KRC to continue conducting valuable ecological research on these data. Consequently, the filenames of the 10-second samples have been randomly sampled, and their temporal order and individual identity cannot be recovered, but can be requested from us.

Fields of Study & Subjects

Music and Audio Processing Animal Vocal Communication and Behavior Speech Recognition and Synthesis

Publication details

Journal / source
Max Planck Digital Library
Publication date
13 March 2026
Language
EN
Document type
Dataset