Both the memory and computational requirements of algorithms traditionally used to extract i-vectors at run time and to train i-vector extractors off-line scale quadratically in the ivector dimensionality. We describe a variational Bayes algorithm for calculating i-vectors exactly which converges in a few iterations and whose computational and memory requirements scale linearly rather than quadratically. For typical i-vector dimensionalities, the computational requirements are slightly greater than those of the traditional algorithm. The run time memory requirement is scarcely greater than that needed to store the eigenvoice basis. Because it is an exact method, the variational Bayes algorithm enables the construction of i-vector extractors of much higher dimensionality than has previously been envisaged. We show that modest gains in speaker verification accuracy (as measured by the 2010 NIST detection cost function) can be achieved using high dimensional i-vectors. Keywords: Text-Independent Speaker Recognition, Features for Speaker Recognition
Cite as: Kenny, P. (2012) A small footprint i-vector extractor. Proc. The Speaker and Language Recognition Workshop (Odyssey 2012), 1-6
@inproceedings{kenny12_odyssey,
author={Patrick Kenny},
title={{A small footprint i-vector extractor}},
year=2012,
booktitle={Proc. The Speaker and Language Recognition Workshop (Odyssey 2012)},
pages={1--6}
}