Speech Enhancement Using Linear Prediction and Dct Coefficients

One implicit assumption the speech enhancement algorithms is that the representation of speech in a transform domain or over a redundant dictionary is sparse, while that of noise is dense. Based on this assumption, clean speech can be recovered by finding the sparse representations. However, some kinds of noise are also found sparse in the above representation scenarios, which results in degradation of enhancement performance. For example, since coefficients of car interior noise are sparse in DCT domain, the speech enhancement performance for car interior noise is not as good as that in other noisy background. In addition, some features, for example, speech energy and the inter-frame correlation, are not considered sufficiently in the available speech enhancement algorithms which probably hinders the further improvement of speech quality. In the proposed method, speech enhancement is casted to an optimization problem, where linear prediction residual and DCT coefficients are combined and adopted as the representation of speech to ensure that noise is dense in such domain. Other features, including speech energy, noise energy and correlation are also considered as constraints to improve the quality and intelligibility of recovered speech. The proposed algorithm adopts LP residual as one of the sparse representation of speech, considering it is feasible and advantageous, as analyzed in the previous section. To make full use of the sparsity of speech, DCT coefficients are also included to contribute as a measurement. The proposed algorithm aims to recover the clean speech, whose LP residual and DCT coefficients are both sparse, via solving an optimization problem under a series of constraints. .