Quasi-Factorial Prior for i-vector Extraction

Abstract

We analyze the i-vector extraction from the perspective of the prior distribution exerted on the mean supervector of Gaussian mixture model (GMM). To this end, we start off with the analysis of the subspace prior which leads to the compressed representation in the standard i-vector extraction. We then propose the use of quasi-factorial prior and show how it impacts the total variability space and its application for i-vector extraction. The quasi-factorial prior could be used in a standalone manner, or in combination with a subspace prior. In the latter context, we found that the performance of the standard i-vector can be greatly improved with the use of quasi-factorial prior followed by a subspace prior. This assertion is confirmed through experiments conducted on the NIST 2010 Speaker Recognition Evaluation (SRE10) dataset.

Incorporating Local Acoustic Variability Information into Short Duration Speaker Verification

“…Supervectors can be regarded as representations of GMMs that differ only in their mixture means [15] and since the total variability model may only provide an incomplete representation for short durations utterances, direct modelling of the supervectors may be beneficial. Parameter tying across mixtures in the total variability model is relaxed in [16,17] and banks of local variability vectors or concatenated local vectors are obtained. G-PLDA was then trained on top them.…”