Oblique Projection and Cepstral Subtraction in Signal Subspace Speech Enhancement for Colored Noise Reduction

Abstract

In this paper, a subspace speech enhancement method handling the case of colored noise using oblique projection in the cepstral domain is proposed. Perceptual features and variance normalization are used to reduce the residual noise and improve the intelligibility of the output speech. Initially, the additive noise present in the noisy speech is removed by removing the orthogonal noise subspace from the noisy speech subspace to obtain the speech subspace. Then, the oblique projection of the noise subspace on the speech subspace along the additive noise subspace is used to determine the colored noise that remains. Colored noise removal is performed by power spectral subtraction in the cepstral domain. The spectral domain constrained estimator that incorporates the combined masking property of the human auditory system is employed to estimate the clean speech signal using the variance of the colored noise. To avoid the occurrence of any abrupt spikes in the output, variance normalization is performed by adaptively changing the control parameter of the estimator's gain matrix. The spectrograms, the objective measures and the subjective intelligibility test show the superior performance of the proposed method over the other existing speech enhancement methods.