Direction finders such as MUSIC experience loss of resolution and increased bias in the presence of nonwhite noise. This paper presents two versions of a steepest descent gradient algorithm that prewhiten the signal received by an arbitrarily oriented volumetric sensor array, minimizing these undesirable effects. The algorithms optimize a whiteness functional over a surface with desirable properties including low dimensionality, unimodality, and concavity. Two ambient noise models facilitate algorithm development through succinct parameterization. The first, which is a novel linear matricial ambient noise model based on a spherical harmonic expansion, places no constraints on array geometry. The second model requires a spatially uniformly sampled sensor array and reduces problem dimensionality associated with exact multidimensional autoregressive modeling. The algorithms estimate a stacked vector arrangement of the model parameters. Application with MUSIC demonstrates enhanced performance in terms of angular resolution and detection of low signal-to-noise ratio (SNR) sources.