Performance evaluation of a real‐time LPC coding technique

Abstract

There are a number of LPC‐based speech synthesis integrated circuits available on the market today. For many practical applications of these chips, it is advantageous to utilize a real‐time code generation system. Traditional LPC analysis techniques, such as covariance, autocorrelation, and PARCOR, process blocks of speech samples yielding one set of LPC parameters for each block. Autocorrelation and covariance techniques require matrix inversion and are too complex for hardware implementation. PARCOR‐type lattice techniques require a large amount of data storage at each stage of the lattice. In this study we have evaluated a fast and efficient LPC analysis technique suitable for hardware implementation for real‐time operation. Unlike the traditional LPC analysis techniques, the real‐time technique processes individual samples of speech. The LPC parameters are first initialized and then updated as each sample is processed. Computer simulation has shown that the real‐time technique yields reasonable spectra. Computer generated synthetic sounds from the real‐time technique have almost the same quality as those generated from the block‐data techniques. Sounds synthesized using the real‐time technique on our SP0256 chip are indistinguishable from those synthesized using traditional techniques.