Speech feature extraction method of improved KPCA

In this paper, we propose a novel speech feature extraction method using kernel principal component analysis (KPCA) based on kernel fuzzy K-means clustering. First, all frames of speech signal are divided into a given amount of clusters by kernel-based fuzzy K-means clustering and then features are extracted by KPCA, as a result of which the storage and computational complexity can be reduced and the original signal can be well represented. Moreover, the proposed method has effects of reducing noise and eliminating tedious information by mapping original eigenvector to a lower dimension space. Simulations show that compared with the existing speech feature extraction methods, the proposed method has a real-time performance, a high speech recognition rate and a better robustness in noisy environment.