Feature-based speaker-independent word recognition without oral learning

This paper describes Multifon, a speaker-independent word recognition system; its originality is based on the fact that Multifon doesn´t require any oral learning; a study realised on phoneme speaker-independent properties allows us to build a permanent dictionnary of phonetic properties for each phoneme of the French language. The recognition process uses this dictionnary to compute the evolution along the time of 29 phonemic distances before a global comparison which realise a time normalisation beetween each word of the vocabulary and the unknown signal.