Fast likelihood computation techniques in nearest-neighbor basedsearch for continuous speech recognition

摘要

This paper describes two effective algorithms that reduce thencomputational complexity of state likelihood computation innmixture-based Gaussian speech recognition systems. We consider anbaseline recognition system that uses nearest-neighbor search andnpartial distance elimination (PDE) to compute state likelihoods. Thenfirst algorithm exploits the high dependence exhibited among subsequentnfeature vectors to predict the best scoring mixture for each state. Thenmethod, termed best mixture prediction (BMP), leads to further speednimprovement in the PDE technique. The second technique, termed featurencomponent reordering (FCR), takes advantage of the variable contributionnlevels made to the final distortion score for each dimension of thenfeature and mean space vectors. The combination of two techniques withnPDE reduces the computational time for likelihood computation by 29.8%nover baseline likelihood computation. The algorithms are shown to yieldnthe same accuracy level without further memory requirements for thenNovember 1992 ARPA Wall Street Journal (WSJ) task