Training of phoneme models in a sentence recognition system

Abstract

This paper describes the training of phoneme models used in a speaker-dependent continuous-speech understanding system. Three different methods for estimating model parameters are described which are based on the standard Markov modelling approach. The first method deals with phoneme models with continuous-emission probability density functions. For the second and third method phoneme models with discrete probability-density functions and two different parameter-estimation methods are described. The test and training speech-database consists of two independent sets of spoken sentences of several speakers. The complete recognition vocabulary contains 917 words with an overlap of 51 words (e.g. articles) with the training vocabulary. Recognition results are given for the different training methods and some other experiments.