14th International Congress of Phonetic Sciences (ICPhS-14)
San Francisco, CA, USA
In recent years, various means of improving the quality of speech
recognition lexicons by augmenting them with pronunciation
variants and phonological rules have been explored.
Alternative pronunciations can be generated using a
Dynamic Programming (DP) approach, whereby different
pronunciations of a word are compiled and any significant
pronunciations extracted. However, such approaches do not
compensate for co-articulation or Hidden Markov Model (HMM)
misclassifications occurring in the observed phone sequence.
Hence, the lexicon may contain poorly defined or spurious
This paper proposes a novel means of automatically compiling pronunciation alternatives, through an approach which removes co-articulation effects and HMM misclassifications. This is accomplished through a probabilistic expansion of the observed pronunciations, using statistical information of how likely the observed pronunciation has been subjected to the aforementioned effects.
The approach, when applied to the TIMIT database was found to improve the quality of the pronunciations generated using a DP based approach.
Bibliographic reference. Stewart, Darryl / Hanna, Philip / Ming, Ji / Smith, F. Jack (1999): "An improved method of automatically generating alternative lexicon pronunciations", In ICPhS-14, 1713-1716.