Analysis of BUT-PT Submission for NIST LRE 2017

In this paper, we summarize our efforts in the NIST LanguageRecognition Evaluations (LRE) 2017 which resulted in systemsproviding very competitive and state-of-the-art performance. Weprovide both the descriptions and the analysis of the systemsthat we included in our submission. We explain our partitioningof the datasets that we were provided by NIST for training anddevelopment, and we follow by describing the features, DNNmodels and classifiers that were used to produce the final systems.After covering the architecture of our submission, weconcentrate on post-evaluation analysis. We compare differentDNN Bottle-Neck features, i-vector systems of different sizesand architectures, different classifiers and we present experimentalresults with data augmentation and with improved architectureof the system based on DNN embeddings. We presentthe performance of the systems in the Fixed condition (whereparticipants are required to use only predefined data sets) andin addition to official NIST LRE17 evaluation set, we also provideresults on our internal development set which can serve asa baseline for other researchers, since all training data are fixedand provided by NIST.