We implemented a PC housed continuous speech acoustical front-end (AFE) for a vocabulary of a thousand words, incorporated into an overall man-machine dialogue system for data base access using telephonic speech. The AFE can be instructed to perform both isolated word and continuous speech recognition; in the latter case the recognition is done in two stages, the first being an evaluation of a lattice of words and of the best scored word sequence, and the second a verification of the most acoustically likely sentences chosen from amongst those which are syntactically and semantically correct. For real-time interactive behaviour, the PC is equipped with 1 DSP board for features extraction and 3 DSP boards with the corresponding memory extension for recognition. Appropriate task synchronization and partition between these boards and the PC program provide real-time performance and a satisfactory degree of accuracy.
Bibliographic reference. Ciaramella, A. / Clementino, D. / Pacifici, R. (1991): "A PC-housed speaker independent large vocabulary continuous telephonic speech recognizer", In EUROSPEECH-1991, 1341-1344.