In 1989, we first reported on the development of summit, a segment-based speaker-independent continuous-speech recognition system . The initial version of summit made use of a small set of context-independent models for the lexical labels. In this paper, we describe our recent attempts to develop a framework that can produce an arbitrarily complex lexical representation. The procedure should permit us to achieve simultaneously the goals of determining a set of context-dependent labels and a lexical network representing alternate pronunciations of the words in our lexicon. Our experiments thus far have been conducted independently on two separate recognition tasks. In both cases, a significant reduction in recognition error rate has been realized.
Bibliographic reference. Phillips, Michael / Glass, James / Zue, Victor W. (1991): "Automatic learning of lexical representations for sub-word unit based speech recognition systems", In EUROSPEECH-1991, 577-580.