| Summary: | 碩士 === 國立交通大學 === 電機與控制工程系 === 91 === This thesis proposed a new text-independent speaker identification technique. A popular decomposition called principal component analysis (PCA) is widely used for feature extraction. PCA has the ability to find out the principal components which are mutually orthogonal. We may also reduce the dimension of feature through these principal components. In the beginning, we construct a time-frequency covariance matrix using the original feature extracted from each frame. Then use PCA to obtain a transformation matrix to get better feature. However, the identification rate is not very good. So we propose some approaches to improve the performance.
At first, we propose another decomposition called independent component analysis (ICA). ICA is more and more popular in recent year because it ability to find out the independent components which are mutually independent. Using ICA, we get an improvement at most 3.61% than using PCA.
Besides, we introduce an optimizer called genetic algorithm (GA). It can provide an optimal solution from all. We want apply GA to reduce the dimension while still maintaining good identification rate. GA optimizer can choose the “best” set of the ICA basis. Using GA optimizer, we get an improvement at most 4.17% than using randomly chosen set of ICA basis; the result is even better than using all PCA basis.
|
|---|
