Frequency warping by dynamic programming

The author introduces dynamic frequency warping (DFW), an exact analog of dynamic time warping (DTW) in which two utterances are represented by matrices, spectrograms that have been sampled in both frequency (perhaps by a discrete Fourier transform) and time (perhaps every centisecond). DFW tests all the dilations in the frequency direction that can be generated by single repetitions of rows of either matrix, and chooses the one that makes the dilated matrices look maximally alike. Thus DFW frequency-aligns similar speech events in the two tokens, as DTW aligns them in time.<
>