Speaker identification using orthogonal and discriminative features

It is desirable that acoustic vectors would form separable clusters in the feature space; however analysis of the common feature vectors does not support this assumption. This paper proposes a new method that manipulates the original features to produce a new feature set in which classes have more convex shape. The proposed methodology uses the idea that according to it, different features have unequal discrimination properties between speakers; So an automatic weighting function based on Multivariate Analysis Of Variance Algorithm (MANOVA) is proposed. MANOVA searches for a linear combination of original features with the largest separation among the speakers. The Vector Quantization (VQ) algorithm is used to detect speakers in the next stage. Although this algorithm is faster and has fewer complexes than the other classification algorithms in this content, promising results are achieved.