Abstract
This paper describes a new method for speech coding called Mixed Excitation Prototype Waveform Interpolation (MEPWI). This method models the speech excitation as the sum of a quasi-periodic deterministic component which evolves smoothly with time and a stochastic component which exhibits unpredictable fluctuations. The deterministic component is represented by interpolation of a sequence of prototype waveforms, while the stochastic component is modeled parametrically as random noise with time-varying frequency content and power level. These two signals are added together to form a single mixed excitation which is used to excite an LPC synthesis filter. When this method is used to model voiced speech, excellent speech quality is ob tained over many kinds of input signals, including breathy speech and speech in acoustic background noise.

This publication has 7 references indexed in Scilit: