This paper describes a framework for optimising the structure and para
meters of a continuous density HMM-based large vocabulary recognition
system using the Maximum Mutual Information Estimation (MMIE) criterio
n. To reduce the computational complexity of the MMIE training algorit
hm, confusable segments of speech are identified and stored as word la
ttices of alternative utterance hypotheses. An iterative mixture split
ting procedure is also employed to adjust the number of mixture compon
ents in each state during training such that the optimal balance betwe
en the number of parameters and the available training data is achieve
d. Experiments are presented on various test sets from the Wall Street
Journal database using up to 66 hours of acoustic training data. Thes
e demonstrate that the use of lattices makes MMIE training practicable
for very complex recognition systems and large training sets. Further
more, the experimental results show that MMIE optimisation of system s
tructure and parameters can yield useful increases in recognition accu
racy. (C) 1997 Elsevier Science B.V.