Computer Comparison by Dynamic Time Warping

A common task for bioacoustic studies is to compare two signals. The underlying question may be “are these two signals the same or are they different?” Alternatively, it may be more general: “how different are these two signals?” Such questions can be successfully addressed through visual inspection of spectrographs, especially using a careful experimental design. However, if a large number of signals are to be compared, visual inspection can require an unmanageable length of time, plus keeping your visual comparisons accurately calibrated is more and more difficult.

Computer-based comparisons of acoustic signals offer the promise of some impressive benefits: the algorithm will not change its assessment over time, it is likely to be much faster than human comparison, and if the algorithm can be accurately described, so can the basis for the comparison. Moreover, as more is discovered about acoustic perception, there is the possibility that algorithms can be tuned using biologically based parameters. However, at the moment such detailed tuning is not really possible, and the first downside of computer-based algorithms is that they reflect the program written by a human being – there is not necessarily a strong link between the comparison algorithm and what organisms actually do. Secondly, anecdotal evidence suggests that no computer algorithm has reached the sophistication of the human visual system for comparing two signals. All of this is to say: computer-based comparisons are no panacea (at least not yet!).

Nevertheless, they are increasingly becoming a valuable weapon in a bioacoustician’s arsenal. Luscinia has an implementation of a fairly advanced comparison algorithm for comparing individual elements, dynamic time warping. It also integrates element comparisons for comparisons of larger units of sound (e.g. songs). Finally, Luscinia incorporates some statistical analysis tools, such as generating UPGMA trees and NMDS graphs of the distances between signals to visualize results.

Having specified a comparison scheme, carrying out a computational comparison in Luscinia involves three steps:

1) Select the acoustic parameters that you wish Luscinia to use, and how they are weighted.
2) Decide which hierarchical levels of organization (element, syllable, syllable transition, and song) you wish to carry out comparisons at.
3) Decide what types of statistical analysis of the results of the comparisons you wish to carry out.