A new method is introduced which tolerates distortions between spectrogram patterns in template matching. Cross-correlation provides a fast method to calculate similarities but it is sensitive to durational and spectral fluctuations. Dynamic time warping tolerates durational differences but requires that frequency components match. An extension to dynamic time warping is introduced where frequency is warped in addition to time. This allows for a detailed analysis to find the similarities and differences between the sound units to be compared. The method was applied to analyses of owl and Bengalese finch vocalizations.
Template matching, dynamic time warping, dynamic frequency warping, pattern recognition