Preserving acoustic cues in speech denoising

Recently methods for obtaining sparse representations of signals from overcomplete dictionaries have been studied and some interesting connections between them and biological sensory processing have been observed. Given that a representation of this type is designed to have only a small number of coefficients different from zero, one might intuitively expect that they would also show a certain robustness to additive noise. In the present paper an example, using speech signals, is examined by comparing several denoising techniques, including a simple method proposed by the authors. The method maintains a sparse representation of the signal and is found to preserve important acoustic cues necessary for phoneme identification.