Multilevel and channel-compensated language recognition: ATVS-UAM systems at NIST LRE 2009

This paper presents the systems submitted by ATVS – Biometric Recognition Group at 2009 language recognition evaluation, organized by the National Institute of Standards and Technology of United States (NIST LRE’09). Apart from the huge size of the databases involved, two main factors turn the evaluation into a very difficult task. First, the number of languages to be recognized was the biggest of all NIST LRE campaigns (23 different target languages). Second, the database conditions were strongly variable, with telephone speech coming from both broadcast news, extracted from Voice Of America (VOA) broadcast system, and conversational telephone speech (CTS). ATVS participation consisted of state-of-the-art acoustic and highlevel systems incorporating session variability compensation via Factor Analysis. Moreover, a novel back-end based on anchor models was used in order to fuse individual systems prior to one-vs.-all calibration via logistic regression. Results both in development and evaluation corpora show the robustness and excellent performance for most of the languages (among them, Iberian languages such as Spanish and Portuguese)