My understanding of Dynamic Time Warping is that the algorithm always requires calculation with each comparison/training series and that there is no way to extract the "essence" from a given training series in the form of coefficients, which could then be compared to coefficients for the test series similar to what might be possible with different wavelet transforms, Discrete Cosine Transform, etc. In experimenting with Python's fastdtw package, DTW does appear to be somewhat slow (despite the name of the package).
What I'm wondering is whether there are any methods that produce coefficients that still preserve the dynamic warping aspect. Perhaps there are certain modifications to wavelet transforms that could do the trick?
