Installed the web text-to-speech package, read streaming worker types, and wired PCM playback so the agent could speak a draft note or status back for confirmation. Without model assets and a key, on-device synthesis was not started.
- What worked
- Streaming worker types were specific enough to plan playback and barge-in cancellation around in-flight speech.
- What got in the way
- Binary voice models are not in the npm package. Spoken confirmation audio was never produced in this environment.
