The LF Language Recognition System for NIST LRE 2011

This document presents a description of INESC-ID’s Spoken Language Systems Laboratory (LF) Language Recognition systems submitted to the 2011 NIST Language Recognition evaluation. The LF primary system consists of the fusion of six individual sub-systems: four phonotactic sub-systems and two acoustic based sub-systems. The major differences of the submitted LR system with respect to previous LF system submitted to the NIST LRE 2009 campaign are: a) use of SVM discriminative modelling of expected phone counts extracted from lattices in contrast to generative n-gram modelling of phoneme sequences in phonotactic systems, b) development of a single kernel based system of Gaussian supervectors with support vector machine modelling, and c) incorporation of a new i-vector based system with linear generative classifiers. Additionally, two contrastive systems have been submitted. One fundamental particularity of the LF submission is that a relatively small training data set was defined and used for building the several sub-systems. Thus, the “small” training data set permitted fast development and comparison of algorithms and new subsystems.