David K. Mellinger
Christopher W. Clark
Bioacoust. Res. Prog., Cornell Univ., 159 Sapsucker Woods Rd., Ithaca, NY 14850
A central problem in bioacoustic transient (animal sound) detection and classification has been the lack of standards against which to compare various methods. A similar problem existed in speech recognition research; it was solved by the creation of databases (e.g., TIMIT) used as common data sources for testing and comparing different methods. A similar database has been initiated for use in marine mammal sound detection and classification. Initially, mysticete sounds have been placed in the database; since their low frequency makes possible a low sampling rate, a large number of sounds may be stored in the space available. The database currently holds sounds of four species of mysticete---1765 blue, 3909 finback, 356 minke, and 589 bowhead whale vocalizations---at sampling rates from 100 Hz to 2 kHz. It was desired to categorize sounds by signal-to-noise ratio, since performance of detection methods will depend on this; since no measure of SNR for transient signals is well known, a simple measure of energy ratio within the sounds' frequency band was used. The database includes performance data of some detection methods, and it is hoped that other researchers will contribute both sounds and performance data. The database is accessible at ftp.ornith.cornell.edu.