The deployment of machine listening algorithms in real-world application scenarios is challenging. In this paper, we investigate how the superposition of multiple sound events within complex sound scenes affects their recognition. As a basis for our research, we introduce the Urban Sound Monitoring (USM) dataset, which is a novel public benchmark dataset for urban sound monitoring tasks. It includes 24,000 sound scenes that are mixed from isolated sounds using different loudness levels, sound polyphony levels, and stereo panorama placements. In a benchmark experiment, we evaluate three deep neural network architectures for sound event tagging (SET) on the USM dataset. In addition to counting the overall number of sounds in a sound scene, we introduce a local sound polyphony measure as well as a temporal and frequency coverage measure of sounds which allow to characterize complex sound scenes. The analysis of these measures confirms that SET performance decreases for higher sound polyphony levels and larger temporal coverage of sounds.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=21683
(683KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
