Out-Of-Vocabulary Detection and Confidence Measures for Speech Recognition Using Phone Models

This paper describes a fast and efficient method to detect out-of-vocabulary words and compute confidence measures in a command-based speech recognition system. The method uses a phone-loop model to reject out-of-vocabulary words and a filler model to compute a confidence measure for each accepted word present in the recognizer output. Tests with this method show that it achieves a good trade-off between falseacceptance versus false-rejection rate. The system runs in real time in a platform with low computational resources and operates in noisy environment conditions (industrial environments and inside vehicles).