Dynamic match phone-lattice searches for very fast and accurate unrestricted vocabulary keyword spotting

The ability to search for typically out-of-vocabulary terms such as names, acronyms and foreign words is a requirement of many audio indexing applications. To date, such applications have employed unrestricted vocabulary keyword spotting approaches that unfortunately suffer from poor miss rates or slow query speeds. This paper proposes a very fast and accurate keyword spotting approach named dynamic match phone-lattice keyword spotting. Reported experiments on conversational telephone speech and microphone speech demonstrate that the proposed method dramatically outperforms conventional methods and is capable of searching at speeds in excess of 300 times real-time while maintaining low miss rate performance.