ParDTW speeds up sequence alignment by 100x on GPUs

A new study from researchers Daniel Yang, Thaxter Shaw, and TJ Tsai introduces four parallelizable alternatives to Dynamic Time Warping (DTW) for aligning long sequences, with a focus on reducing wall clock time by leveraging GPU parallelism. While prior work aimed to lower total computation or memory costs, this approach targets hardware optimization. The algorithms include three approximate methods that break the pairwise cost matrix into rectangular regions, and one exact method that processes the cost matrix along diagonals.

The standout algorithm, Parallelized Diagonal DTW (ParDTW), delivers exact DTW alignments with a runtime reduction of 1.5 to 2 orders of magnitude compared to current alternatives. Another algorithm, weakly-ordered Segmental DTW (WSDTW), provides a faster approximation. All algorithms were tested on an audio-audio alignment task. The paper, published in IEEE/ACM Transactions on Audio, Speech, and Language Processing, offers comprehensive evaluation of accuracy, runtime, and limitations. This breakthrough makes it practical to align very long sequences (e.g., hours of audio) in real time using common GPU hardware.