Large Vocabulary Recognition of Wall Street Journal Sentences at Dragon Systems.

摘要

In this paper we present some of the algorithm improvements that have been made to Dragon's continuous speech recognition and training programs, improvements that have more than halved our error rate on the Resource Management task since the last SLS meeting in February 1991. We also report the 'dry run' results that we have obtained on the 5000-word speaker-dependent Wall Street Journal recognition task, and outline our overall research strategy and plans for the future. In our system, a set of output distributions, known as the set of PELs (phonetic elements), is associated with each phoneme. The HMM for a PIC (phoneme-in-context) is represented as a linear sequence of states, each having an output distribution chosen from the set of PELs for the given phoneme, and a (double exponential) duration distribution. In this paper we report on two methods of acoustic modeling and training. The first method involves generating a set of (unimodal) PELs for a given speaker by clustering the hypothetical frames found in the spectral models for that speaker, and then constructing speaker-dependent PEL sequences to represent each PIC. The 'spectral model' for a PIC is simply the expected value of the sequence of frames that would be generated by the PIC. The second method represents the probability distribution for each parameter in a PEL as a mixture of a fixed set of unimodal components, the mixing weights being estimated using the EM algorithm. In both models we assume that the parameters axe statistically independent.