Acoustic Bird Activity Detection on Real-Field Data

We report on a research effort aiming at the development of an acoustic bird activity detector (ABAD), which plays an important role for automating traditional biodiversity assessment studies – presently performed by human experts. The proposed on-line ABAD is considered an integral part of an automated system for acoustic identification of bird species, which is currently under development. In particular, taking advantage of real-field audio recordings collected in the Hymettus Mountains east of Athens, we investigate the applicability of various machine learning techniques for the needs of our ABAD, which is intended to run on a mobile device. Performance is reported in terms of recognition accuracy on audio-frame level, due to the restrictions imposed by the requirement of run-time decision making with limited memory and energy resources. We report recognition accuracy of approximately 86% on a frame level, which is quite promising and encourages further research efforts in that direction.