TTS

Local text-to-speech via extensions/voicebox/wrapper.py -- thin HTTP client for the voicebox backend. voicebox is not pip-installable: the user clones jamiepine/voicebox and runs python -m backend.main inside the checkout. Models are downloaded by the backend itself.

Commands

W=extensions/voicebox/wrapper.py

python $W doctor                                        # liveness + models
python $W models                                        # list engines/models
python $W generate --text "..." --voice <id> --out out.wav [--engine <e>]

All output is JSON (status, action, result|error, duration_s). Backend down reports status: missing with a start hint; never fabricate audio or a voice list.

Workflow

1. doctor once per session; missing means tell the user to start python -m backend.main in their voicebox checkout.
2. models to pick an available engine/voice; use only listed voices.
3. generate per line or paragraph; keep segments short for easier re-takes. Generation on CPU is slow -- expect tens of seconds per sentence; a GPU-backed engine is strongly recommended for anything beyond short clips.

Narration pipeline

For narrated videos compose across wrappers:

1. tts (this skill) -> per-line WAV files.
2. extensions/asr -> transcribe generated audio for word-level timestamps (subtitle/sync anchor).
3. animation-direction skill -> drive scene/animation timing from the transcript timestamps.

Boundaries

• Localhost only; VB_HOST overrides must be loopback.
• No voice cloning of real people without consent.
• Backend-managed models only; no manual model path injection.