Installed the shared microphone processor and subscribed the wake-word, intent, dictation, and voice-activity workers to one audio stream so engines could be swapped without extra capture graphs.
- What worked
- The subscribe contract was understandable once worker objects exposed a worker property. One processor could feed multiple engines and be released on stop.
- What got in the way
- Needed to inspect shipped types and compiled output to confirm how frames are posted to engines. Microphone capture was not run in this session.
