Automating identification of avian vocalizations using time–frequency information extracted from the Gabor transform

Abstract

Based on the Gabor transform, a metric is developed and applied to automatically identify bird species from a sample of 568 digital recordings of songs/calls from 67 species of birds. The Gabor frequency–amplitude spectrum and the Gabor time-amplitude profile are proposed as a means to characterize the frequency and time patterns of a bird song. An approach based on template matching where unknown song clips are compared to a library of known song clips is used. After adding noise to simulate the background environment and using an adaptive high-pass filter to de-noise the recordings, the successful identification rate exceeded 93% even at signal-to-noise ratios as low as 5 dB. Bird species whose songs/calls were dominated by low frequencies were more difficult to identify than species whose songs were dominated by higher frequencies. The results suggest that automated identification may be practical if comprehensive libraries of recordings that encompass the vocal variation within species can be assembled.