| Summary: | 碩士 === 長庚大學 === 資訊工程研究所 === 94 === In this thesis we use speech recognition technology to recognize rhinitis patients and normal patients. First, we have to build voice database which include rhinitis patients and normal patients. The gender of voice database include female and male. The voice of voice database include /a/ and /m/. Second, we use Mel-Frequency Cepstral Coefficients (MFCC), Formant, Energy, to be different voice features. Third, we use two kinds of classfier. Including Gaussian mixture model (GMM) and Linear Discriminant Analysis (LDA). We use GMM to be acoustic model. LDA is exploited to increase the classification accuracy at a lower dimensional feature vector space. In our experiments, we found MFCC is the best voice feature. Because of MFCC outperforms Formant and Energy. The voice classification accuracy of female /a/ and /m/ are 84.5%. The voice classification accuracy of male /a/ is 69.73% and /m/ is 76.31%. If Lda is applied, MFCC can achieve classification accuracy as follows, The voice classification accuracy of female /a/ is 87.16% and /m/ is 86.8%. The voice classification accuracy of male /a/ is 73.11% and /m/ is 76.83%.
|
|---|
