MAESTRO Synthetic - Multi-Annotator Estimated Strong Labels

The dataset was created for studying estimation of strong labels using crowdsourcing.

It contains 20 synthetic audio files created using Scaper, the reference annotation created with Scaper, and the annotation outcome. Annotation was performed using Amazon Mechanical Turk.

Audio files contain excerpts of recordings uploaded to freesound.org.(from Urban Sound 8k dataset). Please see FREESOUNDCREDITS.txt for an attribution list. 

The dataset contains: 

audio: the 20 synthetic soundscapes, each 3 min long

ground truth:  the "true" reference annotation created using Scaper

estimated strong labels: the reference annotation created from the crowdsourced data

audio tags: the weak labels corresponding to each 10 s segment of the soundscapes, as annotated

For details on the annotation procedure and label processing methodology, see the following paper:

Irene Martin Morato, Manu Harju, and Annamaria Mesaros. Crowdsourcing strong labels for sound event detection. In IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA 2021). New Paltz, NY, Oct 2021.