Dragon Systems and Lernout & Hauspie Speechproducts are jointly developing large-vocabulary speaker-dependent discrete speech-recognition systems in German, Spanish, French, Italian, and Dutch. These systems use the same strategy for training as does the DragonDictate-30K English-language product. Models are created for all the phonemes in the language, each in a wide variety of contexts. The resulting phoneme-in-context models are converted to hidden Markov models whose parameters can be re-estimated on the basis of a modest amount of adaptation data. Prototype systems have now been developed for the five languages. These systems all have vocabularies of a few thousand words and operate on an 80386-based personal computer. The performance of the systems is generally dose to that achieved in English about 85% of the words are recognized correctly, and the majority of errors can be corrected with a single keystroke.
Bibliographic reference. Bamberg, Paul / Demedts, Anne / Elder, John / Huang, Caroline / Ingold, Charles / Mandel, Mark / Manganaro, Linda / Even, Stijn van (1991): "Phoneme-based training for large-vocabulary recognition in six european languages", In EUROSPEECH-1991, 175-182.