A carousel restoration barn wants a talking figure. Official additional material points to ACE with Riva speech, Audio2Face, and a NeMo LLM. What kind of architecture is that stack?

Official additional material that points a carousel restoration barn to ACE with Riva speech, Audio2Face, and a NeMo LLM describes a multimodal speech-and-face avatar architecture selected from listed tools—not a Helm-only cluster fabric exam. Expanding VIA on the spot as a replacement for Audio2Face invents naming and swaps stacks. A text-RAG-only pipeline without speech synthesis or face animation is a different exam story (and drifts toward text-only associate scope). Keep VIA as initials when it appears elsewhere; here the listed ACE/Riva/Audio2Face/NeMo set is the avatar tool-selection answer. Exam caveat: tool-selection from official extra materials only; do not invent expansions or sibling-cert nesting. Operational check: checklist that speech input/output, face animation, and the LLM reply path are all present for the talking figure before calling the stack complete. Checklist ACE speech, Audio2Face, and NeMo LLM presence for the carousel figure before sign-off. Avatar tool-selection uses listed ACE/Riva/Audio2Face/NeMo pieces; do not invent VIA expansions or swap in text-RAG-only stacks. Checklist speech and face paths before calling the carousel figure complete. Close q097 with a municipal evidence note that records why the correct mechanism fits this civic multimodal scenario, why each of the three distractors fails by a distinct concept, and which operational check was observed in logs or dataset cards before promotion. File that note with the run packet so later maintenance cannot silently revive a rejected framing.