A dialogue system based on keyword spotting technology

This paper presented a practical dialogue system, Smart-Panda System, which used a keyword spotting (KWS) scheme. In this KWS scheme, decoding and spotting processes were designed to be independent of the vocabulary and any explicit filler model was not needed so that it can be implemented easily for any other desired tasks. A word level confidence measure was incorporated into this system for improving the accuracy of semantic understanding by rejecting uncompetitive keyword hypothesis, which was typically generated by OOV (out-of-vocabulary) speech. Testing results of both the KWS scheme itself with 100 keywords and the Smart-Panda System, which employed this scheme, were reported. The experiments showed that the correct acceptance accuracy of the Smart-Panda System based on the new KWS recognition engine had reached 95.2% at the incorrect acceptance rate of 0.5%.