Annual competitions are held under the name DCASE (Detection and Classification of Acoustic Scenes and Events), in which researchers all over the world present their methods for the automatic detection of acoustic events (e.g. door slamming, fans, dogs barking) or scenes (e.g. cafeteria, beach) to compare them against each other.
In the DCASE 2022 Challenge Task 4 "Sound event detection in domestic environments" takes Janek Ebbers from the Department of Communications Engineering the first place among 100 submitted systems (find the results here https://dcase.community/challenge2022/task-sound-event-detection-in-domestic-environments-results)!
The task was to develop a classifier that can recognize typical noises that occur in a household (e.g. dishwasher, telephone, vacuum cleaner, etc.). A particular challenge was that most of the training data was not annotated, i.e. had no class labels, and the annotated part of the training data was only weakly annotated. This means that the event class is specified, but not when the acoustic event occurs within the recording. Nevertheless, the system to be developed should output not only the event class but also the timestamp when the event occurred, whereby of course several acoustic events can occur at the same time.
Congratulations to Janek for this great success!
Reinhold Haeb-Umbach
