We present NESTER (Networked Environmental Sonic-Toolkits for Exploratory Research), an open source, scalable, web accessible system that uses advanced machine learning methodologies to analyze large collections of bird recordings. The NESTER system has discovered sex based differences in vocalization patterns, produced accurate syllable transcription, and identified rare flight calls in migrating bird recordings. To use today’s massive computing power to achieve high accuracy, the system uses optimization methods to search for the ideal combination of spectral representation and learning algorithm control parameter settings for each given prediction problem. Parallel processing is employed at both the machine level (multiple cores) and cluster level using cloud computing techniques to fully utilize all available computing resources. The system is implemented in Meander, the parallel data flow language at the heart of the SEASR (Software Environment for the advancement of Scholarly Research). In a typical NESTER application, a signal / no signal model is first built to rapidly parse a recording into interesting and non-interesting segments. The interesting segments then become the focus for the manual tagging of events. Once enough events are tagged, a final model is built to tag events in new recordings.
|Original language||English (US)|
|Title of host publication||Joint Meeting of the Wilson Ornithological Society and Association of Field Ornithologists, 8-12 April 2009, Pittsburgh, Pennsylvania|
|State||Published - 2009|