RECOGNITIONSYSTEMPavelMatˇ.PDFVIP

USE OF ANTI-MODELS TO FURTHER IMPROVE STATE-OF-THE-ART PRLM LANGUAGE RECOGNITION SYSTEM Pavel Matejka,ˇ Petr Schwarz, Luka′sˇ Burget and Jan Cernocˇ ky′ Speech@FIT group, Brno University of Technology, Czech Republic ? matejkap,schwarzp,burget,cernocky ? @fit.vutbr.cz ABSTRACT the following points: This paper concentrates on PRLM (phoneme recognizer followed ￡ using phoneme lattice rather than strings for both training and by language model) approach to language recognition. It elabo- scoring by phonotactic models. This approach was pioneered rates on our prior work concerning the quality of phoneme recog- by LIMSI [4] with good results, and our results with phoneme nition and amounts of training data for phoneme recognizer train- lattices (though our approach was simpler) were also very sat- ing. It reports improvements brought to our PRLM system by better isfactory. phoneme recognition and Witten-Bell discounting in LM-modeling. ￡ use of anti-models: phonotactic models trained on mis- The paper then concentrates on the use of phoneme lattices and recognized segments that should help to discriminate between anti-models. Training and scoring on phoneme lattices brought sig- target and non-target language. Similar approach was used ni?cant improvement in language recognition accuracy. The anti- by SRI in large vocabulary continuous speech recognition models are simple, yet powerful technique to improve the discrim- (LVCSR) [5] to compensate for hypothesis that are acousti- ination between target and non-target languages. All results are cally confusable with the correct transcriptions, we have how- reported on standard NIST 2003 data