Improving the MS-TDNN for word spotting

An MS-TDNN (multistate time delay neural network) version of a word spotting system has been implemented and tested on two spontaneous continuous speech databases. The state-of-the-art MS-TDNN word spotter described here has shown its strength on the RoadRally database. With improvements, it has been adapted to the much more difficult Switchboard corpus with good results. This word spotting system has proved to be a viable alternative to the much larger full-vocabulary speech recognition systems. With relatively few parameters, one is able to achieve good performance and speed on noisy, telephone-quality spontaneous recordings.<>