ADR-0009: Voices and lip sync as provider chains

Context

Rhubarb Lip Sync. v1.14.0 (April 2025) is the latest release and the project is mostly dormant. Its macOS release is Intel-only, and it crashes under Rosetta on this Mac. There is no Linux arm64 build. It builds and runs natively from source on Apple Silicon: CMake, Boost headers only, target rhubarb only. It accepts WAV (8–32-bit, 16–96 kHz) and Ogg, and takes the transcript with -d, which improves accuracy.
macOS say. System voices are licensed for personal, non-commercial use only, so they cannot voice shareable demo content.
Local TTS. Kokoro-82M (code, weights and G2P all Apache-2.0) is the best permissive local option and returns word timings. Caveat: its training data includes synthetic audio from closed models.
Cloud TTS. Several cloud voices return alignment data: phoneme timings from Cartesia and Hume, visemes from Azure, per-character timings from ElevenLabs.

TTS chain, per show: recorded (approved takes, pinned by hash) → configured cloud providers → kokoro (local, via a uv subprocess). Takes are cached by hash(text, voice binding, delivery, provider+version) and stored as FLAC.
Lip-sync chain: rhubarb (built from source by tools/rhubarb/install.sh, pinned to v1.14.0, run with -d <transcript> --extendedShapes GHX -f json) → later an alignment provider that maps TTS or forced-alignment phoneme timings onto the same A–H/X shapes. The engine sees only the resolved viseme track, never the provider.
Demo dialogue for M1 is generated once with Kokoro, committed as FLAC under CC BY 4.0, with provenance recording the model, voice and settings.

No step in the M1 demo depends on a network service or a paid key.
Rhubarb is replaceable without touching the Timeline format.
The toolchain grows by one optional Python environment. It is only needed to regenerate demo audio or voice new lines locally.