This is actually something I'm interested in. It doesn't make sense for real time DSP because the overhead of sending data up to your GPU and back down is too high. However for offline processing it could definitely speed up certain operations (anything done on a per sample basis, which does exclude serially implemented filters. )
The catch is that not everyone has an NVidia GPU, and even those that do may have varying compute versions so it's difficult to write code that works everywhere. OpenCL is an option, and Microsoft has software called AMP that handles this pretty well - we thought about using it for drum replacer but ended up parallelizing on the CPU.
To be honest CPUs are very fast these days. There are definitely audio operations that scale well, but they rarely come up during a normal workflow and are usually interrupted by a third party plugin (that may or may not have a CUDA implementation.) I think it's something that could happen in the future, though - maybe the next generation of PCI buses will be fast enough to support real time DSP.