The GTH-LID System for the Albayzin LRE 12 Evaluation

This paper contains a description of the data-sets, systems and fusion alternatives developed by the Speech Technology Group (GTH) for the Albayzin 2012 Language Recognition Evaluation for the 4 conditions: plenty-closed condition (core), plenty-open, empty-closed, and empty-open. In all cases, the primary system is the fusion of three different i-vector based systems: one acoustic system, a phonotactic system using trigrams of phone-posteriorgram counts, and another acoustic system based on RPLP features instead of the traditional MFCC features. For each plenty condition, a contrastive system was also included where the RPLP features or MFCC features where replaced by a different system based on using glottal source features. We provide results for the plenty conditions using the proposed metrics for the evaluation (i.e. Fact, Fdis, and Fcal), as well the known Cavg metric used on NIST evaluations.