원문정보
초록
영어
In this paper, class-dependent acoustic features are investigated to discriminate consonant class in dysarthric speech. In dysarthric speech, imprecise articulation causes different distortion of consonants depending on each consonant class. For this reason, discrimination of consonant class can play an important role in dysarthric speech processing as a preprocessing step. Therefore, to discriminate each consonant into one of five different classes such as stop, fricative, affricate, nasal, and glide, the discrimination is performed based on a support vector machine (SVM) is employed, which is constructed by using the class-dependent acoustic features combined with mel-frequency cepstral coefficients (MFCCs). It is shown from the discrimination experiments that an SVM using the class-dependent acoustic features relatively reduces average discrimination error rate by 7.67%, compared to that using only MFCCs.
목차
2. Effects of Pronunciation Variations on Consonant Class
3. Proposed Class-Dependent Acoustic Features
3.1. Acoustic Features for Stops
3.2. Acoustic Features for Fricatives
3.3. Acoustic Features for Affricates
3.4. Acoustic Features for Nasals
3.5. Acoustic Features for Glides
4. SVM-Based Consonant Class Discrimination
5. Performance Evaluation
6. Conclusion
Acknowledgements
References