A Novel Instance Matching Based Unsupervised Keyword Spotting System
Peng Li, Jiaen Liang, Bo Xu · 2007
In this paper, we present a novel keyword spotting (KWS) method derived from traditional acoustic KWS. The advantage of this method is that it doesn't need any manually transcribed data to train the acoustic model, so it can be deployed fast for KWS task dealing with small languages and dialectal speech, which the traditional KWS systems can't handle because of the lack of training data. A prototype system is presented, and experimental results shows that this system can achieve very good performance for words that with no less than 3 syllables.