Sixth European Conference on Speech Communication and Technology
(EUROSPEECH'99)

Budapest, Hungary
September 5-9, 1999

Improved Feature Vector Normalization for Noise Robust Connected Speech Recognition

Juha Häkkinen, J. Suontausta, Ramalingam Hariharan, M. Vasilache, K. Laurila

Nokia Research Center, Speech and Audio Systems Laboratory, Tampere, Finland

Feature vector normalization has been successfully usedto improve the noise robustness of speech recognizers.Unfortunately, it may cause additional insertion errors inconnected digit recognition in clean environments. Wepropose two methods to reduce the number of insertions.Based on estimated instantaneous signal-to-noise ratiowe form a reliability measure for the recognized digits.We discard unreliable digits from the beginning and theend of the recognized digit sequence. Since the proposedreliability hypotheses are independent of the likelihoodsproduced by an HMM classifier, we are capable ofbringing new useful information into the classificationprocess. In addition, we constrain the normalizationprocess on the basis of statistics obtained from thetraining data. Experimental results show that we arecapable of achieving an average 32% string level errorrate reduction in simulations of a noisy car environment.


Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Häkkinen, Juha / Suontausta, J. / Hariharan, Ramalingam / Vasilache, M. / Laurila, K. (1999): "Improved feature vector normalization for noise robust connected speech recognition", In EUROSPEECH'99, 2833-2836.