On The Use Of Convolutional Neural Networks In Pairwise Language Recognition

Convolutional deep neural networks (CDNNs) have been successfully applied to different tasks within the machine learning field, and, in particular, to speech, speaker and language recognition. In this work, we have applied them to pair-wise language recognition tasks. The proposed systems have been evaluated on challenging pairs of languages from NIST LRE' 09 dataset. Results have been compared with two spectral systems based on Factor Analysis and Total Variability (i-vector) strategies, respectively. Moreover, a simple fusion of the developed approaches and the reference systems has been performed. Some individual and fusion systems outperform the reference systems, obtaining similar to 17% of relative improvement in terms of minC(DET) for one of the challenging pairs.