4pAB6. Spectrogram contour features for intraspecific classification of narrow-band animal sounds.

Session: Thursday Afternoon, May 16

Time: 2:50

Author: Ben Pinkowski
Location: Dept. of Comput. Sci., Western Michigan Univ., Kalamazoo, MI 49008


Recent research has indicated that contour features obtained from computer imaging are useful for characterizing some interspecific differences in narrow-band animal sounds [B. Pinkowski, J. Acoust. Soc. Am. 95, 3419--3423 (1994)]. Contour features may also characterize the intraspecific differences of animal sounds that vary among individuals. To examine this possibility, six songs from each of ten eastern bluebirds (Sialia sialis) (5 males and 5 females) recorded during 1991 and 1992 in southwestern Michigan were digitized at 22 kHz, converted to spectrogram images, and analyzed for contour using Fourier descriptors (FD's). Following smoothing, thresholding, and segmentation, the spectrogram components corresponding to three or more basic notes were merged by a simple pixel aggregation algorithm to yield a single contour per song. The contour border was then used to compute a 512-point DFT. With 16 low-order FD's per song, 58 of the 60 songs (96.7%) were correctly classified by a linear discriminant function. Male songs were more varied and complex and hence more difficult to classify by contour alone than were female songs. [Work supported by NIH.]

from ASA 131st Meeting, Indianapolis, May 1996