In this paper, a new matching method by fuzzy inference is proposed. In vowel, variations of the formant frequencies depend on speakers and spoken situation. It is difficult to decide a prior value of width of the variations, so we represent their variations as fuzziness and make fuzzy rules. For consonant recognition, it is difficult to extract feature parameters. Therefore, here we use characteristics about configuration of autocorrelation function, power, zero crossing rates, and logarithmic energies.
As a result of experiment on the recognition of continuous Korean speech, it is shown that the fuzzy inference is effective to speech recognition of unspecified persons.
抄録全体を表示