Combining DCT and Adaptive KLT for Noisy Speech Enhancement

摘要

This paper investigates the correlation between successive speech components across time as well as frequency in DCT domain, and proposes a novel speech model for enhancing noisy speech, which assumes the sequence of speech components among successive frames to be a highly correlated and non- stationary process. Based on this model, a linear estimator of clean speech components is obtained from the noisy speech components of successive frames using MMSE estimation. This estimator is implemented by applying the KLT to the noisy components vector. And to reduce the computation demand, an adaptive KLT technique is used for performing eigenvalue decomposition in agreement with the speech model. In simulations with speech signals degraded by noises, the proposed method shows improved performance over the traditional DCT method for a number of objective and subjective measures.