Method for recognizing speech using linguistically-motivated hidden Markov models

A method for modeling word-initial acoustic cross-word effects for use in generating word pronunciation networks from a set of pronunciation rules and a training database on a data processing apparatus, wherein said training database comprises representations of speech signals of words being pronounced in a continuous speech manner, said speech signal representations being linked to textual representations of said words, wherein word pronunciation networks include textual representations of words and corresponding phonetic networks, said modeling method comprising the steps, for every word network w, of:

• for every initial arc a.sub.i, setting a variable p.sub.i equal to a phone label on said arc a.sub.i ;
• and
• for every phone label p.sub.j in a phonetic inventory,
• a) counting a number of occurrences c in said training database of the word w preceded by any word ending with said phone p.sub.j ;
• thereafter
• b) if c is greater than a preselected threshold,
• i) adding an initial arc a.sub.k to said word network w with phone label p.sub.i which connects to a common "to-node"[as said arc a.sub.i ;
• thereafter
• ii) constraining said arc a.sub.k to only connect to arcs with label p.sub.j ;
• and
• iii) constraining said arc a.sub.i not to connect to arcs having phone label p.sub.j.