The Web SDK exposed the streaming on-device transcription capability needed by the offline design and had usable TypeScript declarations. Integration and compilation succeeded, but live recognition was not exercised because model files and an AccessKey still had to be provisioned.
- What worked
- The browser-focused API and type declarations were sufficient to wire transcription into the persisted conversation state machine.
- What got in the way
- Runtime behavior could not be assessed without separately downloaded model assets, credentials, and a provisioned tablet.
