13th Annual Conference of the International Speech Communication Association

Portland, OR, USA
September 9-13, 2012

Phonotactic Language Recognition Using MLP Features

Mohamed Faouzi BenZeghiba, Jean-Luc Gauvain, Lori Lamel

Spoken Language Processing Group, LIMSI-CNRS, Orsay, France

PPRLM language recognition systems using context-dependent (CD) phone recognizers outperform significantly those using contextindependent (CI) phone recognizers, but computationally are less efficient. This papers describes a very efficient PPRLM system both in terms of performances and processing speed. The system uses CI phone recognizers trained with MLP features concatenated with the conventional PLP and pitch features. MLP features have some interesting properties that make them suitable to build such a system, in particular the temporal context provided to the inputs of the MLP and the discriminative criterion used to learn MLP parameters. Results of preliminary experiments conducted on the NIST LRE 2005 for closed-set task show significant improvements (for the three conditions) obtained by the proposed system compared to a PPRLM system using CI phone models trained with PLP features. More ever, the proposed system performs equally compared to the PPRLM using CD phone models, while running 6 times faster.

Index Terms: Language recognition, Phonotactic approach, MLP features

Full Paper

Bibliographic reference.  BenZeghiba, Mohamed Faouzi / Gauvain, Jean-Luc / Lamel, Lori (2012): "Phonotactic language recognition using MLP features", In INTERSPEECH-2012, 2041-2044.