In this paper, we investigate Multilingual Stacked Bottleneck Features (SBN) in language recognition domain. These features are extracted using bottleneck neural networks trained on data from multiple languages. Previous results have shown benefits of multilingual training of SBN feature extractor for speech recognition. Here we focus on its impact on language recognition. We present results obtained with monolingual and multilingual networks, and their fusions. Using multilingual features, we obtain 16% relative improvement on 3 s condition of NIST LRE09 dataset with respect to features trained on a single language.
Bibliographic reference. Fér, Radek / Matějka, Pavel / Grézl, František / Plchot, Oldřich / Černocký, Jan (2015): "Multilingual bottleneck features for language recognition", In INTERSPEECH-2015, 389-393.