UPS-IRIT system #2 for track1: kmeans+select

We run a k-means (k=100 clusters, L2 norm) on the baseline MFCCs (static, first and second derivative, 39-d) that are ZCA whitened on a per-file basis. We then re-estimate the centroids after selecting the data points for which the assignment label is the same as the one of its left and right nearest neighbors.

The feature representations correspond to the distances between the data points and the cluster centroids.