Second European Conference on Speech Communication and Technology

Genova, Italy
September 24-26, 1991


Bottom-Up Acoustic-Phonetic Decoding for the Selection of Word Cohorts from a Large Vocabulary

Henri Meloni, F. Bechet, P. Gilles

Laboratoire d'Informatique, Faculté des sciences, Avignon, France

Within the framework of the analytical recognition of the words of a large vocabulary, we propose a system permitting the selection of limited cohorts of lexical items from a lattice of valued phonetic units. Before the system can work, it has to learn all the phonetic units of a given speaker by storing the spectra of each phoneme. During the construction of the phonetic lattice, the units are simultaneously localized and identified by means of various types of spectral distances adjusted according to the phonemes, the context and certain characteristics of the speaker's voice. A few factors dependent on the phonemes as well as various types of spectral stability make it possible to determine the kind of sound and the areas in which the scores will be calculated. In the first stage the system interprets the information available in the lattice and then it accesses the lexis through a series of filters thus reducing the number of possible words. The solutions are given in the form of a cohort of words that are assigned a value according to the scores and the rate of temporal covering-up of the phonemes identified.

Full Paper

Bibliographic reference.  Meloni, Henri / Bechet, F. / Gilles, P. (1991): "Bottom-up acoustic-phonetic decoding for the selection of word cohorts from a large vocabulary", In EUROSPEECH-1991, 647-650.