Articulatory description of speech signal in isolated word recognizer

A method of speech signal description to be used in moderate-size vocabulary word recognizer is presented. This method is based on the assumption that in a task-specific system each word pronounced in isolation can be represented as a sequence of articulatory elements, not necessarily of phoneme size, which would be sufficient for its recognition. A general description of speech signal based on characterization of the manner of articulation was applied using context-independent algorithms transforming each word into a sequence of following, discrete articulatory units: resonants, stridents, other obstruents and unvoiced closure. In the process of mapping acoustic cues into articulatory features fuzzy relations were used. For each 10 ms segment of the speech signal the plausibility of a given type of articulation was evaluated, using fuzzy characterization of time evolutions of broad-band spectral energies. In the next step, some specific, short sequences of these (2 or 3) units are transformed into larger segments corresponding to more complex, transitional articulatory movements, connected with production of some speech sounds (such as apical /r/) or some characteristic, underlying articulatory events (namely bursts). Such approach has given satisfactory results with few ambiguities. Sequence of different types of units forms an articulatory skeleton of the analysed word and underline its matching with the vocabulary. Some experimental results obtained for 6 speakers are presented.