Implementing the dynamic time warping algorithm in multithreaded environments for real time and unsupervised pattern discovery

Dynamic Time Warping (DTW) has been a widely used algorithm in the field of patter recognition. DTW is used to finding acoustic similarities in the same speech sequence or between sequences or both. Its use is not limited to speech signals but it is also a key step in image processing as well. Despite being one of the most important and effective algorithms, DTW is computationally very intense. Processing of one hour of speech using DTW takes a few hours on a single processor, limiting its applicability to desktop and server platforms. Even on advanced platforms, DTW is used only in an offline manner and not in real time. Further modifications for improving performance in DTW make the algorithm slower. In this paper, we aim at extracting maximum thread-level parallelism from the process so as to accelerate its execution using clusters, multicore and multi-processor servers. Since the existing parallelism in this process in highly limited, we restructure the entire algorithm to extract maximum parallelism without altering the functional behavior of the algorithm. We implement the algorithm on a cluster of Intel Xeon processors running at 2.93GHz. We compare the results on a multi processor and multicore level to analyze the benefits of both versions. Our results show that it is possible to implement such a compute intense algorithm in real time which is a big boost considering that these algorithms are always done in an offline manner.