Spherical harmonics really are the "spherical Fourier transform" you're looking for. The kind of hack you mention in comments, of doing a 2D Fourier transform on a lat-long projection, suffers from all the problems you usually have when you try to project a sphere onto a plane: not all spatial relations in the sphere are well-represented in the plane. If you take stretching at the poles as an example, that clearly results in a frequency shift. High-frequency components in the sphere get shifted down to low frequencies in the plane, and when you do the inverse transform, any inaccuracy shows up as high-frequency components at the pole.
Such a transform is useless for analysing the spectrum of the original signal, because it distorts the frequencies (and distorts them differently in different parts of the sphere). It's also poor for representation, because of the introduction of these high-frequency artefacts.
Spherical harmonics use different basis functions so they can work directly on the lat-long without treating it like a plane. If you plot the response of the spherical harmonics on a plane projection of the sphere, they look kooky (like a map of the earth does), but they look much more natural on a sphere. They don't suffer from artefacts at the poles, and they don't frequency-shift signals. They're also rotation-invariant, so it doesn't even matter where you put the poles. They don't even have poles, in the sense of discontinuities, which is a property you can't achieve from a plane projection however many ad hoc co-ordinate transforms you do. Rotation-invariance is useful for applications like environment maps, where you'd like to rotate the spherical-harmonic transform of the environment (light probe, or whatever) into the local co-ordinate system before feeding it to the shader. There isn't a projection onto the plane that has this property.
