We are currently studying how prosodic cues could facilitate continuous speech recognition and understanding. In this paper, results are discussed concerning the detection of syntagm boundaries from the analysis of fundamental frequency and rhythm variations. Three corpus have been tested corresponding to differents speech production conditions: reading, memorized sentences (short-term memory), quasi-spontaneous information dialogues. Our present rates for the detection of word or syntagm boundaries are roughly: 90% correct markers.
Bibliographic reference. Carbonell, Noelle (1989): "On the use of prosodic knowledge for continuous speech recognition and understanding", In EUROSPEECH-1989, 1522-1525.