Abstract: Based on a large number of speech experiments, Mandarin speech recognition approaches have been thoroughly studied, and a real-time speaker-dependent all-syllable recognition system of Mandarin has been developed on an IBM PC/AT microcomputer with a high-speed digital signal processing board TMS320C25-E. In accordance with the phonetic characteristics of Mandarin, the three-stage recognition strategy is adopted in this system. Experiments for the speech datas of 4 times 1240 syllables show that, average correct rate of four tone recognition is about 99%, correct rates of the first 5 candidates of syllable recognition are 82%, 91%, 94%, 96%, and 97% respectively, and the whole system response time is less than 0.2 second. In addition, the Mandarin initials and finals confusion matrices, and the corresponding hierarchical clustering diagram of the similarity are obtained from the experiment results, and they are analyzed in comparision with the references 1,2 so as to further improve the system performance.