The relaxation of a dissipative system to its equilibrium state often shows a multiexponential pattern with relaxation rates, which are typically considered to be independent of the initial condition. The rates follow from the spectrum of a Hermitian operator obtained by a similarity transformation of the initial Fokker-Planck operator. However, some initial conditions are mapped by this similarity transformation to functions which grow at infinity. These cannot be expanded in terms of the eigenfunctions of an Hermitian operator, and show different relaxation patterns. Considering the exactly solvable examples of Gaussian and generalized Lévy Ornstein-Uhlenbeck processes (OUPs) we show that the relaxation rates belong to the Hermitian spectrum only if the initial condition belongs to the domain of attraction of the stable distribution defining the noise. While for an ordinary OUP initial conditions leading to non-spectral relaxation can be considered exotic, for generalized OUPs driven by Lévy noise these are the rule. The relaxation of a physical system, prepared in a nonequilibrium state, to the equilibrium often shows a multiexponential pattern with decrements of single exponentials defining the relaxation rates. These rates are usually considered an intrinsic property of the system, independent of initial conditions and they follow from the spectrum of a Hermitian Hamiltonian operator obtained by a similarity transformation of the Fokker-Planck (FP) operator governing the evolution of the probability density. Methods of spectral analysis are central in physics, in particular in quantum mechanics and in the theory of oscillations, and are universally employed to the solution of linear problems. Thus the discussion of the spectrum of the FP operator is often the first step in the solution of the FP equation and the investigation of its relaxation properties [1-3]. As we proceed to show, this first step might not deliver a complete picture. Initial distributions which are not mapped to square integrable functions by the similarity transformation, cannot be expanded in terms of the eigenfunctions of the corresponding Hamiltonian operator and will therefore relax at rates that may not be given by the Hermitian spectrum. It is in this sense that we use the term non-spectral relaxation. The smallest nonspectral rate can be smaller than the smallest spectral relaxation rate and thus dominate the relaxation behavior over the whole time range. Although the effect of non-spectral relaxation can be observed under quite general conditions, in this Letter we concentrate on simplest, exactly solvable examples of Ornstein-Uhlenbeck processes (OUPs) describing the coordinate of an overdamped particle in a harmonic potential driven by a white noise. Because the OUP generally approximates random processes in the vicinity of a stable stationary point it is a very important analytical tool in many fields of research, from statistical physics [3][4][5], to theoretical neuroscience [6], ecology [7] and economics [8]. Since the ass...