- Phonetically sensitive transformation of speech features yields
improvement in performance
- Two-level HMM structure is used - one for describing the
sentence in terms of words and another for describing words in
terms of phones
- Observation models for each state are assumed to be
multivariate guassian
- Phonetic discriminant model transforms each speech feature
vector by a state-specific linear transformation
- Reference models are word-level models with two models for each
word - one for men and one for women