Supervector Compression Strategies to Speed up I-Vector System Development

Abstract

The front-end factor analysis (FEFA), an extension of principal component analysis (PPCA) tailored to be used with Gaussian mixture models (GMMs), is currently the prevalent approach to extract compact utterance-level features (i-vectors) for automatic speaker verification (ASV) systems. Little research has been conducted comparing FEFA to the conventional PPCA applied to maximum a posteriori (MAP) adapted GMM supervectors. We study several alternative methods, including PPCA, factor analysis (FA), and two supervised approaches, supervised PPCA (SPPCA) and the recently proposed probabilistic partial least squares (PPLS), to compress MAP-adapted GMM supervectors. The resulting i-vectors are used in ASV tasks with a probabilistic linear discriminant analysis (PLDA) back-end. We experiment on two different datasets, on the telephone condition of NIST SRE 2010 and on the recent VoxCeleb corpus collected from YouTube videos containing celebrity interviews recorded in various acoustical and technical conditions. The results suggest that, in terms of ASV accuracy, the supervector compression approaches are on a par with FEFA. The supervised approaches did not result in improved performance. In comparison to FEFA, we obtained more than hundred-fold (100x) speedups in the total variability model (TVM) training using the PPCA and FA supervector compression approaches.

Who Do I Sound like? Showcasing Speaker Recognition Technology by Youtube Voice Search

“…The UBM is a 1024-component Gaussian mixture model (GMM) [10], which is used to compute sufficient statistics for ivector extraction. We compute 800-dimensional i-vectors by compressing mean supervectors of maximum a posteriori (MAP) adapted GMMs using probabilistic principal component analysis (PPCA) as described in [5]. This is a (speed-wise) high-performing alternative to the stardard i-vector extraction that is traditionally done via front-end factor analysis [11,12].…”

“…The i-vector extraction using PPCA is simply a matter of compressing 61440-dimensional GMM-supervector to 800-dimensional space using a precomputed projection matrix. Note that the traditional approach for i-vector extraction would, in addition, require inverting an 800 × 800 posterior covariance matrix [14,5].…”

“…We run our demo on a web platform that can be used on PCs and mobile phones with an internet connection to ensure good accessibility of the demo. The web platform communicates with a computation server that runs the speaker recognition back end based on our recent work on computationally efficient i-vector extraction [5]. The back end provides the results to the web platform that displays them by using embedded YouTube video players.…”