The IRIT-UPS system @ ZeroSpeech 2017
Abstract
In this report, we describe the IRIT-UPS approach for the unsupervised discovery of sublexical units in speech in the framework of the Zero Resource Speech Challenge 2017 edition. We derive unsupervised representations that consist of the distances between the MFCC vectors extracted from the speech signal and a hundred cluster centroids estimated on millions of these vectors with an efficient scalable implementation of the k-means algorithm. We show that using a whitening transformation (ZCA) to pre-process the MFCCs is crucial to outperform the baseline MFCC representation. Furthermore, we obtain slight but consistent improvements by performing data selection before estimating the centroids. This simple approach yielded competitive results for the within-speaker condition but generalized less well in the betweenspeaker one.