eduzhai > Applied Sciences > Engineering >

Sound Event Detection Using Duration Robust Loss Function

  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: Many methods of sound event detection (SED) based on machine learning regarda segmented time frame as one data sample to model training. However, the sounddurations of sound events vary greatly depending on the sound event class,e.g., the sound event ``fan has a long time duration, while the sound event``mouse clicking is instantaneous. The difference in the time durationbetween sound event classes thus causes a serious data imbalance problem inSED. In this paper, we propose a method for SED using a duration robust lossfunction, which can focus model training on sound events of short duration. Inthe proposed method, we focus on a relationship between the duration of thesound event and the ease difficulty of model training. In particular, manysound events of long duration (e.g., sound event ``fan ) are stationarysounds, which have less variation in their acoustic features and their modeltraining is easy. Meanwhile, some sound events of short duration (e.g., soundevent ``object impact ) have more than one audio pattern, such as attack,decay, and release parts. We thus apply a class-wise reweighting to thebinary-cross entropy loss function depending on the ease difficulty of modeltraining. Evaluation experiments conducted using TUT Sound Events 2016 2017 andTUT Acoustic Scenes 2016 datasets show that the proposed method respectivelyimproves the detection performance of sound events by 3.15 and 4.37 percentagepoints in macro- and micro-Fscores compared with a conventional method usingthe binary-cross entropy loss function.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×