Analysis and recognition of Korean isolated vowels using formant frequency.

Abstract

This paper describes analysis and recognition of eight Korean isolated vowels based on the distributions of formant frequencies. In the distributions of vowels on the F1-F2 plane, we found overlappings between the vowels /u/, /o/, /∂/ and /w/, /ε/ and /e/ overlapped heavily as if they are the same vowels. In the comparative investigation of the dialect of Seoul and Kyungsangdo, Kyungsangdo dialect showed a heavy overlapping between /∂/ and /w/. As a recognition experiment, the speaker-independent recognition with Bayes decision method was carried out for 1, 760 isolated vowels spoken by 20 speakers. A recognition rate of 81.2% correct was obtained with the formant frequencies F1 and F2while the rate with LPC cepstrum coefficients was only 76.4%. Normalization for eliminating inter-speaker differences showed good results for vowels uttered by speakers of the same dialect. The highest recognition rate of 99.8 % was obtained using a supervised learning method.