Sixth European Conference on Speech Communication and Technology
We propose general representation of knowledge and hypotheses used in speech recognition systems. General abstract entity is described and further specialised to represent various types of information - time intervals, speech signal, sequences of Markov model states, pronunciation, written text etc. These types of data represent tiers of an overall knowledge about the utterance and during the recognition process not all the tiers are fully known. Furthermore, mutual alignment of individual tiers is not fully known. Operations with these multitier partially known structures have common definition on the abstract level and individual datatypes are created by further restrictions of the basic abstraction. Common properties are however strong enough to enable design of algorithms optimizing the whole recognition system.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. Hanzl, Václav (1999): "Theory of structured cogitation in speech recognition", In EUROSPEECH'99, 1247-1250.