the find
− 0.6 mA digital twin that can hold a live voice call, hands-free, in Thai or English. It listens, thinks, speaks in a cloned voice, and lip-syncs onto real recorded footage of its subject — not a rendered head, actual video with the mouth re-inpainted in real time. Everything runs on a single M5 Max laptop with the network cable optional. First contact with the outside world: a live call on 2026-07-27.
cross-section
− 1.3 mThe pipeline is four organs, each swappable between local and remote by a single config file. Speech-to-text is Whisper large-v3 on MLX — 0.37 seconds, faster than the RTX 5090 sitting one room away. The brain is a local LLM served over MLX. The voice is a cloned model on the same machine. The face is MuseTalk on Apple's GPU, inpainting mouth frames onto real footage at 26–30 fps; a diffusion renderer on the desktop tower exists as an optional remote fallback, but the twin currently lives entirely on the laptop. Audio reaches the call through a pair of virtual cables — a topology where one wrong routing makes the caller hear themselves echoed, which is exactly the kind of thing you only learn once.
Fully warm, the twin makes its first sound 5.8 seconds after you stop talking. All local.
inscription
− 2.0 mField note, filed under FACE_RENDERER_RESEARCH: Before committing to a face, a multi-agent survey screened 84 candidate renderers, deep-read 18 of them with a second agent assigned to adversarially verify every claim, and concluded that on Apple Silicon the incumbent still wins — because inpainting a real recording means "the liveliness is genuinely his rather than simulated." The survey also surfaced the silent-mouthing landmine: one source clip had constant mouth movement and no audio, silently defeating loudness-based frame ranking. The fix was to stop trusting the audio and rank frames visually, by jaw motion.