linear prediction

A technique used in SPEECH SYNTHESIS and SPEECH RECOGNITION to represent ACOUSTIC PHONETIC knowledge in a way which is capable of computational processing. In linear prediction coefficient (LPC) synthesis, a speech signal is defined by a set of coefficients (predictors), which try to predict the signal from its past time domain values. These coefficients are then used to produce a REPRESENTATION of the SPECTRUM of the signal. The approach is based on the analysis of RESONANCES in the VOCAL TRACT, and is thus especially useful in its ability to identify FORMANT locations (though sounds involving NOISE features are less accurately modelled), producing syntheses of high quality. The technique is increasingly common in spectral analysis within PHONETICS.