When users respond to automatic speech recognition systems, they usually utter not only desired words but also unexpected ones. To deal with such responses, the recognizer must be able to retrieve valid words and reject invalid ones. This capability, which is called word spotting, is therefore becoming increasingly important as speech recognition systems continue to move from the laboratory to actual applications.
In this paper we propose a post processing algorithm to improve word spotting performance by reducing false alarms while maintaining high recognition accuracy. Compared with the previous approaches, this algorithm increase a little amount of calculations and can be implemented without any modification of conventional parameters. The concept of the algorithm is to rescore the words hypothesized by the primary word spotter using the post processor, which is designed to discriminate between true keyword occurrences and false alarms. The post processor uses weighted likelihood scores of the hypothesized words, and the weights are calculated based on distances between the states of HMM of the words.
Experimental results showed that the word spotting system using the proposed post processor significantly outperforms a baseline system. When the proposed post processor was incorporated into a word spotting system, the false alarm rate decreased by 36.6% when setting the rejection rate to 7.0%.