DCASE 2018 Task 5: Multi-Channel Audio for Domestic Activity Recognition

The DCASE 2018 Task 5 development dataset is a derivative of the SINS database, containing audio for detecting daily activities in a home. It consists of 72,984 ten-second audio segments from four microphone arrays, totaling approximately 200 hours of data. The segments are manually annotated for activities like cooking, dishwashing, eating, and social activity.

Use Cases

• Training audio classifiers for domestic activity recognition based on multi-channel acoustic features.
• Benchmarking multi-microphone array processing algorithms for sound source localization in a home.
• Developing models for presence/absence detection in smart home environments using continuous audio recordings.
• Researching acoustic sensor networks for long-term, unobtrusive activity monitoring.

Strengths

• Contains 72,984 annotated audio segments, providing a substantial corpus for model training.
• Features multi-channel audio data from four microphone arrays, enabling spatial acoustic analysis.
• Uses manual annotation for activity labels, which likely ensures high-quality ground truth.

Limitations

• Metadata completeness is low; specific column names, file formats, and exact update dates are not provided across sources.
• The dataset scale is limited to one person in a single vacation home over one week, which may affect generalizability.
• Access is governed by an End-User License Agreement (EULA) whose specific terms are not detailed in the provided metadata.

Provenance

Source

KU Leuven (Gert Dekkers et al.), derived from the SINS database.

Collection Method

Continuous recording in a vacation home using an acoustic sensor network, later segmented and manually annotated.

Time Range

One week of continuous recording (specific dates not provided).

Geography

A vacation home environment (location not specified).