Third International Conference on Spoken Language Processing (ICSLP 94)
The aim of this paper is to describe a novel application of the Discriminative Training (DT) procedure for non-keyword rejection based on the principle of minimizing a weighted error criterion. This technique is applied to our Keyword Recognition System, a speaker-independent semicontinuous Density Hidden Markov Model (SCDHMM) recognizer. The proposed algorithm is evaluated for two different isolated word recognition tasks on telephone-line recordings containing both keywords and non-keywords utterances. We will compare the results with those obtained with Maximum Likelihood Estimation (MLE). In the Rejection Application the proposed procedure offers an automatic way of tunning the parameters for the desired application and a decrease on the Cost Function. Good results have been also obtained with the Discriminative Trained Garbage Model in the Word- Spotting Application.
Bibliographic reference. Torre, Celinda de la / Acero, Alejandro (1994): "Discriminative training of garbage model for non-vocabulary utterance rejection", In ICSLP-1994, 475-478.