Abstract
This project integrates different parts of a speaker-dependent isolated word voice-activated typewriter on a personal computer (IBM PC-AT). To build up the language model (for French), several routines have been written: automatic grapheme to phoneme conversion, semiautomatic training texts (20 pages) processing (building up the graphemic (2500 words) and phonemic (2000 words) lexicons), syntactic labelling through inductive inference, computation of the probabilistic language model (bigrams and trigrams on grammatical classes), and the definition of the phonological rules. The speech signal is analysed by 20 digital bandpass filters. Several types of speech compression techniques have been tried on medium and large difficulty vocabularies. Vector quantisation and nonlinear time compression have been chosen. Recognition is conducted in three steps: fast match based on word length and gross comparison; detailed match based on conventional DTW algorithms; and use of the language model to take into account the linguistic constraints, and to achieve the phoneme to grapheme conversion. Overall recognition rates of 95% have been obtained with a mean recognition time of 2 s, the 2000 templates being stored on 60 KBytes of RAM memory. Recognition results with or without the language model have been compared. >
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have