Database of Annotated Real Environmental Sounds
DARES is an initiative to stimulate environmental sound perception research by providing a set of annotated databases and associated tools. All contributed databases will be accessible with a single set of tools. Additionally, DARES aims to form a community of researchers interested in environmental and ecological sound recognition. Therefore we aim to maintain an inventory of researchers working with the data.
DARES is initiated by Tjeerd Andringa and Dirkjan Krijnders from the Auditory Cognition Group (ACG) of the Department of Artificial Intelligence of the University of Groningen and in collaboration with INCAS3.
DARES provides two databases:
A 2 hour (1.3 GB) database compiled by Maarten van Grootel, Dirkjan krijnders and Tjeerd Andringa, which was presented on the NAG-DAGA meeting in Rotterdam (see research). This database is augmented with a set of MATLAB tools to search the database and to add or change information.
A 1 hour (200 MB) database compiled by Dirkjan Krijnders in the CASSANDRA project. This database was first presented at the Advanced Video and Signal based Surveillance 2007 Conference(see research).
Additionally, DARES aspires to provide in due time:
- Annotated databases of environmental sounds
- These databases will be open and free for research and educational purposes. Commercial use is encouraged as well, but for a reasonable fee to maintain and improve DARES activities and to reimburse the database developers.
- Environmental sound analysis tools
- DARES will provide MATLAB-tools to work with the XML annotation data, for example search-tools to select all sound samples of a specific class. Furthermore, DARES will provide an annotation tool that is optimized for the demands posed by environmental sound research. The Auditory Cognition Group is working on a tool.
- Performance measures
- One of the problems with real-world input is that it is very difficult to measure performance: human annotations may vary considerably (when is a "far car" far enough to annotated as such?), and multiple answers may be correct (e.g. when do groups of passing cars become traffic?). Furthermore the tenth car in a minute is probably less important than the single car in a natural reserve. The development of good performance measures should become a priority for the research community. DARES aims to provide standardized performance measures.
- Related to the previous point, it might be good to provide results from perceptual experiments
- For modeling purposes and benchmarking it is important to have access to human responses to everyday sounds. For example it may be important to know the typicality and familiarity of sounds samples to interpret recognition results. DARES hopes to provide this type of information as well.
- Forum / mailing list
- When DARES develops sufficient activities it might be worthwhile to start a discussion forum and/or mailing list to help to form a more active community.
- Links to related information and to research groups interested
Plans and expectations
- Currently two more annotated databases are in preparation. One contains five 3-minute recordings at five different locations in the city of Assen (The Netherlands). The other is a 1 hour database of simulated (verbal) aggression related events at a train station in Amsterdam.
- The Auditory Cognition Group has developed an annotation tool. This tool will be made available in due time. This tool will provide annotations in a format that is directly suitable for DARES, and that is readily accessible with the available tools.