Feature request: capturing both microphone input and output

[whitequark]

Once again useful for getting transcript of meetings, including what I'm saying in it.

Running the model twice in parallel is probably impractical, so I think mixing the two inputs would work fine.