Installed the web SDK as on-device dictation and as the fallback when a custom intent context was unavailable. Endpoint-aware transcripts fed a local grammar so commands still worked without Rhino.
- What worked
- Worker types made endpoint detection and transcript accumulation straightforward. The engine could stay unloaded until listening started, which helped the memory budget.
- What got in the way
- The public parameter file is large and the first CDN URL returned 403, so model fetch needed a retry against raw hosting. Live dictation was not exercised without a microphone and AccessKey.
