http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
박창호,안종영,김주성,김수훈,이영재,허강인 동아대학교 공과대학 부설 한국자원개발연구소 1995 硏究報告 Vol.19 No.2
A conventional CHMM is unable to express time transient variation well. In this paper, we formed segments to combine several frames and made up a vector for each segment to reflect a dynamic variation speech. If we used as they were, the number of parameter's dimensions increased. Therefore we could not estimate aptly when the train date was insufficient. So we compressed the parameter's dimensions and decreased the number of parameter's dimensions. At the exteriment, we compared the result of using vectors compressing mel-cepstrum by the K-L transformation with the result of using mel-cepstrum and mel-cepstrum coefficients. At the result, we got higher recognition rates of using mel-cepstrum and mel-cepstrum + regressive coefficients than that of using vectors compressing mel-cepstrum by the K-L transformation.
HMM을 이용한 연속 음성인식 시스템의 화자적응화에 관한 연구
심장엽,김상범,김주성,김수훈,이영재,이종진,허강인 동아대학교 공과대학 부설 한국자원개발연구소 1995 硏究報告 Vol.19 No.2
It is hard to collect sufficient speech data for training a speaker-dependent (SD) model from the same speaker. In contrast, to trains a speaker-independent (SI) model need not collect a large amount of speech data per speaker but from many speakers. Speaker-adaptation (SA) is an additional training technique from SI model by a small amount of adaptation speech. It has proved to be a powerful tool for achieving good recognition performance without the high cost of SD training. In this study, a speaker adaptation algorithm (MAPE) which trains it by every utterance sequentially without hand-labelling is introduces. The hand-labelling is performed automatically by Concatenation training and Viterbi-segmentation. The secuential-training is performed by MAPE(Maximum A Posteriori probability Estimation). We can train it using any small amount of adaptation speech data. For newspaper editorial continuous speech, the recognition rates of adaptation of HMM was 62.5% respectively which is approximately 32.5% improvement over that of unadapated HMM.
안종영,박창호,김상범,김주성,김수훈,허강인 동아대학교 공과대학 부설 한국자원개발연구소 1995 硏究報告 Vol.19 No.2
A feed forward neural network has been used for the pattern classification. It has the capability of representing a desired input-output mapping through the training of a given set of teaching patterns. In this paper, we proposed recognition by GPFN and PNN a kind of RBFN. The neural network can approximate a posteriori probability through Bay's theorm, by training it with binary vectors as target pattern corresponding to the categroies of input pattern. In the phoneme recognition, we compared the recognition rate of GPFN, PNN with Hybrid(VQ, LVQ) Algorithm. We found that recognition rate of LVQ-PNN, VQ-PNN had higher then that of them.