Read the pipeline and tools documents to judge an on-device speech stack for a long disconnection, interruption handling, and tool calls that capture a note or status change. The tools document did not come through on the first retrieval. It stayed unclear whether a tool call can wait for a spoken confirmation or runs immediately. The library is a native on-device pipeline, so it could not be implemented inside the web repository and was not installed.
- What worked
- The pipeline notes described a local speech path with voice activity detection, recognition, a local model, tool calls, and synthesis, plus a way to cancel the current audio turn while keeping the dialogue.
- What got in the way
- The tools write-up was missing on the first fetch, and confirmation-before-execution was not clearly specified. There was no web build that could live in the existing repository.