単音節音声の実時間認識装置

抄録

A voice recognition system is proposed for the input device of Japanese word processor, for the voice control of industrial robots and for the communication aids of the disabled. Our device is operated in real time by user's voices which are pronounced as a series of Japanese open monosyllables. The hardware is designed by use of a microprocessor system so that the total system may be a small size and low cost. Input voice is divided into 16 components by 15 channel switched capacitor filters and an envelope detector, and the most suitable area of the consonant part is extracted by investigating the square distance between the input envelope and the reference one. Every extracted pattern is represented by 16x16 dimensional vector after the logarithmic conversion, the level normalizing and the time smoothing. Each input pattern is compared with the reference monosyllabic patterns of a talker by use of the arithmetric processing unit which is developed for the high speed calculation of the square distance between two voice patterns. The time required for the identification of a monosyllable is about 200 msec. From the experimental results by 67x100 monosyllables uttered by 5 speakers, the average rate of correct identifications shows about 96% and the envelope matching method is proved to be effective especially for the identification of the voiced consonants. However, monosyllables following vowel /i/ or /u/ show 2〜4% lowerer correct rate than others and most of confusions occurs between the unvoiced consonants.