Gaussian density tree structure in a multi-Gaussian HMM-based speech recognition system

This paper presents a Gaussian density tree structure usage which enables a computational cost reduction without a significant degradation of recognition performances, during a continuous speech recognition process. The Gaussian tree structure is built from successive Gaussian density merging. Each node of the tree is associated with a Gaussian density, and the actual HMM densities are associated to the leaves. We propose then a criterion to decide whether a node belonging to a high level in the tree should be expanded or not. The expansion means that the likelihood is evaluated with Gaussian densities associated with a low level node if the likelihood computed at the high level is not precise enough. This Gaussian tree structure is evaluated with a continuous speech recognition system on a telephone database. The expansion criterion allows a 75 to 85% computational cost reduction in terms of log-likelihood computations without any significant word error rate increase during the recognition process.