Minicomputer simulation was carried out to design an optimum system based on currently available signal processing LSI's. First, finite‐word‐length effects of Levinson‐Durbin (LD) algorithm and Le‐Roux (LR) algorithm [J. Le‐Roux et al., IEEE Trans. Acoust. Speech Signal Process. ASSP‐25, 257–259 (1977)] for extracting PARCOR coefficients were investigated regarding (1) the PARCOR/AR/cepstrum coefficient error, (2) the difference in LPC cepstrum distance between the top two candidates, and (3) recognition rate. LR was found to be almost always better than LD by each of above measures. Second, the effects order of analysis, number of template bits, and the template normalization method were examined to minimize memory size. It was shown that the number of template bits of each cepstrum coefficient can be reduced to four with little decrease of recognition rate as compared to the system with floating point number templates. A single‐boarded recognizer using TMS320 for LPC analysis and MN 1263 for DP matching was implemented. The overall recognition rate of on‐line test in speaker‐dependent mode was 99.4% for a ten‐word vocabulary (total of 1000 tokens of ten speakers). Multiple template speaker‐independent mode achieved a recognition rate of 97.0% for an eight‐word vocabulary.