Feature pruning in likelihood evaluation of HMM-based speech recognition

In this work, we present a simple yet effective technique to reduce the likelihood computation in ASR systems that use continuous density HMMs. In a variety of speech recognition tasks, likelihood evaluation accounts for a significant portion of the total computational load. Our proposed method, under certain conditions, only evaluates the component likelihoods of certain features; and approximates those of the remaining (pruned) features by prediction. We investigate two feature clustering approaches associated with our pruning technique. While a simple sequential clustering works remarkably well, a data-driven approach performs even better in its attempt to save computation while maintaining baseline recognition accuracy. With the second approach, we can speed up the likelihood evaluation by 33% and reduce its power consumption by 27% for an isolated word recognition task. For a continuous speech recognition system using either monophone or triphone models, the speedup and power reduction of the likelihood evaluation are 50% and 35% respectively.