Dynamic Time Warping (DTW)

The DTW is a method used in text-dependent Speaker Recognition. In this context, the training or testing data are composed by a sequence of acoustic vectors and the temporal order of the vectors is important. In order to compute likelihood or a distance between two of such sequences, two functions are needed, a frame to frame distance function and a frame mapping function, able to align the individual acoustic frames of both sequences. This time alignment function is mandatory as two occurrences of the same linguistic messages, pronounced or not by the same speaker, present different time characteristics, like the global pronunciation speed. If there is a training template T _r with $N_{T_R }$ frames and a test utterance T _E consisting in a sequence of $N_{T_E }$ frames, the DTW is able to find the time mapping function w(n) between T _R and T _E . In the figure 1, the tying function w(n) is illustrated by the tying of the T _R frame at time x with the T _E frame at time y. Thus, the...