Task independent wordspotting using decision tree based allophone clustering

Two areas of research have been addressed. The first area is the potential viability of task-independent (TIND) training in wordspotting. It has been demonstrated that good task independent performance can be obtained using large, phonetically rich speech corpora for training triphone subword acoustic models. The effect of TIND training corpus size on task-dependent (TDEP) performance has been demonstrated. The second area of research was motivated by the fundamental limitation of TIND performance that exists when using fixed acoustic units. It was found that for even very large TIND corpora, the coverage of TDEP triphones will always be limited, thus demonstrating the need for a technique for learning a suitable acoustic subword unit for the particular task. The use of allophone decision trees for identifying acoustic units in a TIND wordspotter training scenario has been investigated. Preliminary experiments have demonstrated that, by expanding keywords using allophones obtained from a TIND allophone clustering procedure, significant improvement in TDEP wordspotting performance can be obtained.<>