You could always offload some layers to the NPU for lower power use and leave the rest to the GPU. If the latter is power throttled (common for prefill, not for decode) that will be a performance improvement.

That seems like a really niche use case, and probably not worth the surface area? The power savings would have to be truly astonishing to justify it, given what a small fraction of compute time your average device spends processing voice input. I'd wager the 90th percentile siri/ok google/whatever user issues less than 10 voice queries per day. How much power can they use running on normal hardware and how much could it possibly matter?