The Pre-image Problem and Kernel PCA for Speech Enhancement

In this paper, we use kernel principal component analysis (kPCA) for speech enhancement. To synthesize the de-noised audio signal we rely on an iterative pre-image method. In order to gain better understanding about the pre-image step we performed experiments with different pre-image methods, first on synthetic data and then on audio data. The results of these experiments led to a reduction of artifacts in the original speech enhancement method, tested on speech corrupted by additive white Gaussian noise at several SNR levels. The evaluation with perceptually motivated quality measures confirms the improvement.