A k-Nearest-Neighbour Search under the Dynamic Time Warping Metric is often in the literature reported to achieve the highest accuracies.
However, the runtime costs are quite high, so an efficient implementation is key.
I compared different setups and implementations that can be used from Python. This repository contains the best combination that I came up with. It is based on an enhanced DTW C implementation and the kNN algorithm from sklearn which is running parallel.
It is only tested for python 2.7 so far.