Installed the web SDK and used its voice-activity worker to stop spoken prompts when the user talks over them, keeping barge-in on-device and CPU-only.
- What worked
- Binding readme and type definitions were enough to create a worker, subscribe it to the shared audio processor, and run inference on CPU. The small footprint matched the tablet constraints.
- What got in the way
- Barge-in was not verified with a live microphone. Option types needed a close read to confirm a process-error callback could be used without extra device fields.
