Dynamic time warping for transient noise classification

Dynamic time warping (DTW) is a method of defining the similarity of waveforms. In contrast to cross-correlation, DTW allows signals to be stretched and compressed ("warped") in order to find the best match possible. Once a metric of similarity has been defined, a set of many signals can be partitioned into groups that are mutually similar. Separating signals into similar groups can be a useful first step in finding potential sources of noise. In this paper, the theory of dynamic time warping is reviewed with examples. DTW was used to partition recordings of transient percussion sounds into groups, with each group representing a source instrument. After partitioning, an "average" signal for each group was calculated, which was then used to compare the different sources of sound.