초록

AI-Helper

LPC열로 구성된 음성패턴의 주파수변동을 해결하기위해 LPC와 스펙트럼, LPC차수와 스펙트럼의 관계를 고찰하여 새로운 형태의 멤버쉽함수를 제안하였다. 또한, 시간변동을 해결하기위해서는 음성구간을 여러구간으로 등간격분할하는 다구간 등분할법을 사용하였으며, 이때 오인식은 주로 동일음절이 같은 발성위치에 있을때 발생되었다. 이러한 오인식을 줄이기위해 제안된 멤버쉽함수로 퍼지추론한뒤 구간별 확신도에 가중치를 부여하고, 세번째후보까지를 인식대상으로 하는 판정알고리즘을 제안하였다. 본 방법의 타당성을 검증하기위해, DDD지역명 28개를 대상으로 인식실험한결과, 삼각형멤버쉽함수에 의한 퍼지추론은 $92.0\%$, 삼각형멤버쉽함수에의한 퍼지추론과 판정알고리즘은 $92.9\%$, 제안된 멤버쉽함수에의한 퍼지추론과 판정알고리즘은 $93.8\%$의 인식률을 보였다.

Abstract ▼ AI-Helper

To solve the frequency variation of speech patterns which consist of LPC sequences, new membership function view from LPC, spectrum and the relations between the order of LPC and spectrum is proposed. To solve the time variation, multi-secation equi-segmentation method which equally divide the speec...

To solve the frequency variation of speech patterns which consist of LPC sequences, new membership function view from LPC, spectrum and the relations between the order of LPC and spectrum is proposed. To solve the time variation, multi-secation equi-segmentation method which equally divide the speech section into several section are applied. False recognition mainly occur at time when the same syllable is placed at the same utterance. To reduce the error, fuzzy inference is executed using the proposed membership function and weights are assigned into sectional certainty and then the decision method for recognized the section up to the third candidate. To testify the validation of this method, we experimented the recognition test of 28 DDD area names. The recognition rate of the fuzzy inference by the triangle membership function is $92\%$. That of the combined method of the fuzzy inference and the dicision method is $92.9\%$ and that of fuzzy inference by the proposed membership funtion is $93.8\%$.