Convolutional Recurrent Neural Networks for Rare Sound Event Detection
Emre Çakır and Tuomas Virtanen
Tampere University of Technology, Finland



In this page, you can find the illustration of DCASE2017 rare sound event detection challenge evaluation set signals and the target event detection labels for different methods. In the figures, the onset and offset for each target event is labeled for CNN and CRNN methods. If the figure of a sample from the evaluation set cannot be found in this page, it means that all of the methods agree that there are no target events present in the corresponding sample.

You can reach the scientific paper on the proposed method here.

Figure 1: CRNN system overview.

Dataset download links

Dataset is available for download through DCASE2017 official website.
  • (1/5)
  • (2/5)
  • (3/5)
  • (4/5)
  • (5/5)

    Annotated figures of the samples

    In each figure, the estimated onset and offset for the given sound event is annotated for both the CNN and the CRNN methods in the paper.
  • Baby cry (263 figures)
  • Glass break (252 figures)
  • Gun shot (258 figures)


    The annotated figures from each sound event combined with a 5 second delay in a .gif.
  • Baby cry (gif)
  • Glass break (gif)
  • Gun shot (gif)
  • All three classes (gif)