Deriving optimal codes for speech and natural sounds using generalized independent component analysis

Abstract

A Bayesian method for inferring an optimal basis is applied to the problem of finding efficient codes for natural sounds. The key to the algorithm is multivariate non-Gaussian density estimation, which is an equivalent independent component analysis when the form of the marginal density is fixed. An important advantage of the probabilistic framework is that it provides a method for comparing the coding efficiency of different bases objectively, and the derived codes can be shown to have better coding efficiency compared to traditional Fourier and wavelet bases. It also provides a method for Bayesian signal denoising and filling in of missing samples using a basis that is optimized to the structures in the data. When this framework is applied to deriving efficient codes of speech and natural sounds, the codes share many of the coding properties of the cochlear nerve. Time–frequency analysis shows that it is possible to derive both Fourier-like and wavelet-like representations by deriving efficient codes for different classes of natural sounds.