Analyse de la parole par les méthodes de modélisation paramétrique

This paper provides a survey of recent speech analysis techniques. After stressing the value of a precise and accurate analysis technique in most of speech processing applications, the basic parameter extraction methods are critically summarized. Attention is then focused on parametric modelling methods as applied to speech analysis. After a restatement of the linear prediction principles and associated fast algorithms, the three following topics are examined in some depth : ? global analysis methods on short time windows, with variable frequency resolution, and with additive noise; ? global analysis by multipulse techniques ; ? time evolving methods where a time varying parametric model is adjusted to model the transitions between quasistationnary periods; ? time adaptive sequential methods using fast (Kaiman, Cholesky…) algorithms along with a synchronous detection of temporal events.