Two-Stage Isolated-Word Recognition System Using Discriminant Analysis

摘要

This report describes a two-stage isolated-word recognition system using a Hidden Markov Model (HMM) recognizer in the first stage, and a statistical discriminator in the second stage. The second-stage system performs pairwise discriminations between the top few candidate word models when no clear decision is made from the first stage. Likelihood-ratio comparisons and a new technique called sifting are used to focus attention on those features that best differentiate word pairs. This system alleviates four fundamental problems which are found with most conventional speech recognition systems. These problems include: 1) the effects of limited training data are not explicitly taken into account, 2) the correlation between adjacent observation frames is incorrectly modeled, 3) durations of acoustic events are poorly modeled, and 4) features which might be important in discriminating only among specific work pairs or sets of words are not easily incorporated into the system without degrading overall performance. The system was tested on a 35 word/10,000 token stressed-speech isolated-word data base created at Lincoln Laboratory. The adding of the second-stage discriminating system reduced the error rate by more than a factor of 2. The overall error rate fell from 7.7 percent with only the HMM system to 3.5 percent with both the HMM system and the discriminator.