An evaluation of keyword spotting performance utilizing false alarm rejection based on prosodic information

In this paper, we describe our effort in developing new method of false alarm rejection for keyword spotting type of speech recognition system. This false alarm rejection uses prosodic similarities, and works as posterior rescore basis. In keyword spotting, there is always false alarm problem. Here, we propose a technique to reject those false alarms using prosodic features. In Japanese, prosodic information is expressed in intonation form, while may of other languages is using stress accents. Therefore, it is easy to calculate prosodic information using fundamental frequency, so called F0, in our language. In our new keyword spotting engine, we get result by combining two scores. One is phonetic score calculated by front engine, and the other is pitch score calculated by post engine described in this paper. We have accomplished 13%(point) improvement on keyword recognition accuracy using this method. We also have proposed robust modeling method for rejection using prosodic features.