This paper describes the principles and the implementation of a large size lexicon to be used within a continuous speech recognition system, the HEAD system. The lexicon contains the system application vocabulary and it is a major design criterion that a new vocabulary is entered into the HEAD system at lexicon level only. As the lower levels of the system remains unaltered, no additional training of the system is required when the vocabulary is changed. The lexicon is based on partial allophone information realized as a set of distinctive phonetic features. A lookup consists of a search- and a verification- phase. The search phase is based a number of predefined key archiphones, whereas the verification phase utilizes all available information.
Bibliographic reference. Larsen, Lars Bo / Baekgaard, Anders / Bundgaard, Michael (1989): "Constructing a large size lexicon for a continuous speech recognition system", In EUROSPEECH-1989, 2232-2235.