
A voice instruction can be convenient without being a safe motion command. CatForge Voice Lab makes the boundary visible: AssemblyAI Streaming STT v3 produces a formatted Mandarin or English final turn, and the browser places that turn in an editable command box. The operator reviews the request before planning, validation or execution. The deterministic bilingual planner can compose only five reviewed MicroCat primitives: bow, wave, stretch, tail wag and stand. Typed contracts bound duration, amplitude and ordering. A reference gate checks schema and motion limits; a rejected proposal can be revised before the operator explicitly chooses Run. Cancellation, validation results and execution events remain in one ordered trail. An accepted skill exports as a portable ZIP with the plan, evidence, SHA-256 manifest and a Python replay verifier. Another process can check the files and recompute the reference gate without trusting a screenshot. The web console and FastAPI backend provide a complete typed fixture workflow. AssemblyAI uses short-lived browser tokens, while the provider credential stays on the server. Protected voice sessions use an access code and a cumulative budget ledger. Three authenticated provider sessions verified the Streaming v3 lifecycle with generated Mandarin, English and mixed-language PCM audio; recognition errors are retained as correction examples.
30 Sep 2026