First European Conference on Speech Communication and Technology

Paris, France
September 27-29, 1989

Vowel Recognition in a Data Base of Continuous Speech: Experiments with Local and Global Identification Principles

M. J. Caraty, J. C. Richard, Xavier Rodet

Université Pierre et Marie Curie, France

In continuous speech, a vocalic segment presents an evolution resulting mainly of coarticulation effects. The idea of a unique point in the vocalic space, interpretable in terms of speech production, is attractive for the identification method. However, assumption of existence and unicity of such a point in vocalic segments is contradicted by many observations. Consequently, an improvement of recognition rate is expected if the evolution within the segment is taken into account in the identification process. Therefore, we propose a vowel recognition experiment by pattern matching (ie location and identification) on a BDSON (1) data base, including about 1500 vowels, in order to compare the advantages of a global principle of identification (ie on the whole segment) with the advantages of its local counterpart (ie on selected frame(s)).

Full Paper

Bibliographic reference.  Caraty, M. J. / Richard, J. C. / Rodet, Xavier (1989): "Vowel recognition in a data base of continuous speech: experiments with local and global identification principles", In EUROSPEECH-1989, 2272.