How Does a Dictation Machine Recognize Continuous Speech?

by T.Dutoit, L.Couvreur, H.Bourlard

The Chapter 4 of the book "Applied signal processing: a MATLAB-based proof of concept" shows how GMMs are used for the classification of (supposedly stationary) signals and how HMMs provide a means of modelling non stationary signals as sequences of stationary states.
The chapter also implements a simple bigram model, whose coupling with the word HMMs has resulted in a unique sentence HMM, able to perform continuous speech recognition, i.e., to find how many words are present in an incoming stream of feature vectors and what those words are.

Applied signal processing : A MatlabTM-based proof of concept
Thierry Dutoit, Ferran Marques
Springer, 2009