Speech decomposition using the discrete cosine transform, and adaptive thresholding

In this paper we present a new algorithm, named DCT-HN, for decomposing speech into two parts - a harmonic and a noise component. The technique uses adaptive thresholding in the Discrete Cosine Transform (DCT) domain to determine the noise level for each frame. The technique is iterative in nature but converges quickly to a local optimum. The method can be applied to either speech or residual signals. Experimental results for both synthetic and real speech signals showed that DCT-HN yields the same results as a similar decomposition method, referred to as YAD, for high levels of Signal to Noise Ratio (SNR). However, the proposed technique outperforms the most available methods for low SNRs.