To my knowledge, Software Occlusion Culling (which you already mentioned), is pretty much the only thing a software rasterizer would still be used for. Procworld makes use of a similar technique to display its huge voxel environments.
Most other culling methods like frustum culling do work on the CPU, but, to stick with the example, the test against the frustum happens on object level, probably with an axis-aligned bounding box (AABB). This intersection test is way simpler than using a full-blown software rasterizer.
In certain cases, software rasterization could be used for mouse picking of objects. In game engines, this is often solved using a physics engine and ray-triangle collision with a simplified mesh.
With the CPU being idle while waiting for the GPU in modern interactive 3D applications, one could think that it might be beneficial to use these idle cycles to render on the CPU using a software rasterizer. The problem here, besides the rendering getting horribly complex and convoluted, will often be the bandwidth. Images rendered on the CPU need to be transferred to the GPU before getting displayed, which might cancel out the benefit.