Audioset Scripts

A collection of scripts to analyze, prepare for and download Google's Audioset.

audioset_download.py

This file contains a function to download, format and segment a given YouTube audio, as well as a function to process an entire list of files in a parallelized way.

The expected input is a file listing file segments to be downloaded line by line in the format YTID_STARTMS, where YTID is the YouTube-Id of the video and STARTMS is the start time (in ms), from which a 10s interval will be extracted (see train_list.txt and eval_list.txt).

This requires external packages yt-dlp and sox.

audioset_scripts.py

This file contains scripts to counts files, classes and events in the dataset, select top most occuring classes, filter the dataset by a list of files or classes, as well as make tables of counts for several cases.

These scripts assume the usage of Google's Audioset: Reformatted. The files of the dataset need to be placed into src/ folder.

The only external files that are needed are train_list.txt and eval_list.txt, which list the actually downloaded files.