Variable threshold vector quantization for reduced continuous density likelihood computation in speech recognition

Vector quantization (VQ) has been explored in the past as a means of achieving reductions in likelihood computation for hidden Markov models (HMMs) which use Gaussian mixtures for their output densities. In this paper, we present a new method for choosing which mixtures can be discarded for each pair of HMM state and vector quantization index. Traditionally, a global threshold was used to specify the maximum distance a mixture mean could lie from a VQ codeword before being considered negligible in likelihood calculations for observation vectors contained in that VQ cell. Our technique uses a threshold which varies with VQ cell volume. Thus, larger cells are allocated more mixtures than smaller cells, in order to provide a more uniform coverage of the acoustic space and thereby improve computational efficiency.