Task Independent Speech Verification Using SB-MVE Trained Phone Models

Robust ASR-systems should benefit from detecting when portions of the decoded hypotheses are incorrect. This can be done by including a separate verification module based on statistical hypothesis testing. String based minimum verification error (SB-MVE) training is a promising alternative for improving the corresponding verification-models. This paper adresses a variant of SB-MVE at the phone level for design of task independent verification modules. The algorithm updates both H0 and H1 phone models. Experiments are performed on ”time of day” recordings of the Norwegian part of Speechdat (II). The results show a relative decrease in utterance error rate (compared to no verification) from 8 − 37% for false rejection rates ranging from 0 − 25%. Thus the method shows robustness with respect to choice of treshold.