A real-time, talker-dependent, connected-speech recognizer has been operational at the Labaratory for Engineering Man-Machine Systems (LEMS) for 3 years. This recognizer analyses strings of connected digits or alphadigits, using dynamic programming (DP) techniques and an expert for final decision. DP often misclassifies within difjicult subgroups of the vocabulary such as the E-set (letters e, p, t, b, d, g, v, z, c). In this paper, we present a feedback mechanism for disambiguation of words

more »

... ssified to be in the E-set by the first pass of the recognizer. This mechanism reanalyses the speech data within an appropriate context and uses the analysis as input to a hidden Markov model (HMM) which uses acoustic, phonetic and linguistic knowledge about the elements of the Eset. Experiments have been run using speech from 5 male talkers. A different model was computed for each talker. Each model was trained with 12 replications of each word and tested with 72 utterances. A recognition rate of95% was achived.