DASED: A Multi-Domain Dataset for Sound Event Detection Domain Adaptation

Abstract

In this paper we present the first freely available dataset for the development and evaluation of domain adaptation methods, for the sound event detection task. The dataset contains 40 log mel-band energies extracted from 100 different synthetic sound event tracks, with additive noise from nine different acoustic scenes (from indoor, outdoor, and vehicle environments), mixed at six different sound-to-noise ratios, SNRs, (from -12 to -27 dB with a step of -3 dB), and totaling to 5400 (9 * 100 * 6) sound files and a total length of 30 564 minutes. We provide the dataset as is, the code to re-create the dataset and remix the sound event tracks and the acoustic scenes with different SNRs, and a baseline method that tests the adaptation performance with the proposed dataset and establishes some first results.

Tasks

Reproductions