Deep Learning for

bioacoustics

Problem FOrmulation

  • We investigate how to identify bird individuals in field recordings obtained by the Bioacoustics lab in the University of Alberta.
  • We attempt to identify bird individuals in field recordings, paying attention to:
      • Number of individuals per species.
      • Location of the individual from the recording microphone.

Our Current Approach

  • We first train a deep neural network on audio files containing bird songs, collected by the Bioacustic lab of the University of Alberta.
  • The classifier network is convolutional and takes a spectrogram of bird songs as the input.
  • We are testing different parameters used to build the spectrograms used as inputs to determine which help the network classify with higher accuracy. The parameters are:
      • Length of time window.
      • Number of frequency filters.
      • Overlap between time windows.

Spectrogram of the call of a Yellow Warbler