Phonematic recognition by linear prediction: Experiment

The recognition of speech signals analyzed by linear prediction is introduced. The principle of the channel adapted vocoder (CAV) is outlined. The learning of each channel model and adaptation to the speaker are discussed. A method stemming from the canonical analysis of correlations is given. This allows, starting with the CAV of one speaker, the calculation of that of another. The projection function is learned from a series of key words pronounced by both speakers. The reconstruction of phonemes can be explained by recognition factors arising from the vocoder. Automata associated with the channels are used for local smoothing and series of segments are treated in order to produce a phonemic lattice.