Abstract: The INESC-ID's Spoken Language Systems Laboratory (L$^2$F) Language Verification system submitted to the 2009 NIST Language Recognition evaluation is introduced in this paper. Then, as a sequence of the evaluation workshop and post-analysis of the results, the set of modifications that lead to significant performance gains are reported. Main differences between the original submitted system and the post-evaluation system consist of: 1) the kind and amount of training and development data considered for language model training and calibration and fusion, 2) the improvement of the acoustic based sub-systems and the reduction of the number of sub-systems that compose the whole system, and 3) the application of a better calibration and fusion scheme. Contrastive results of the submitted and the post-evaluation language recognition system for the different conditions in the evaluation are provided.
Index Terms: language recognition.