Brag Voice

brag-voice — Brent's voice, locally, free

The voice: Voicebox app (jamiepine/voicebox) on this Mac, API http://127.0.0.1:17493. Profile "Brent Bryson" 18836009-0527-4b9f-9837-3d58c2a0dc76 — cloned from his own footage, approved with chills.

The one-command lane (narrate a render)

cd ~/Desktop/The\ Brag-Machine/machine
./scripts/vo.sh runs/<run-id> "<narration text>" [delay-ms]   # → runs/<run-id>/out-vo.mp4

Generates the speech in Brent's voice, ducks the music to 0.35, mixes, never re-renders video. Copy result to Desktop under a UNIQUE name and open for his review.

Raw speech (no video)

curl -s -X POST http://127.0.0.1:17493/generate -H "Content-Type: application/json" \
  -d '{"text":"...","profile_id":"18836009-0527-4b9f-9837-3d58c2a0dc76","language":"en"}'   # returns {id}
# poll: curl -s "http://127.0.0.1:17493/history?limit=10" | jq '.items[] | select(.id=="<id>") | .status'
# fetch: curl -s "http://127.0.0.1:17493/audio/<id>" -o out.wav

Gotchas (all hit and solved 2026-07-23 — do not rediscover)

• App must be running: open -a Voicebox; wait for API 200 on /profiles.
• /history wraps results in .items[]. /generate/{id}/status is an SSE stream — do NOT curl it with a timeout and parse as JSON; poll /history.
• Server is single-lane: wait for no generating/loading_model items before POSTing (vo.sh does this).
• "librosa stub" error = stale server bundle → quit + relaunch the app.
• New voice samples: max 30s audio + exact reference_text transcript (transcribe with npx hyperframes transcribe; its JSON is an ARRAY of {start,end,text}).
• VOICE DRIFT (2026-07-23, Chrissy video): ellipses/long pauses in the text can make the clone DRIFT to a different voice (even female) mid-read. Write continuous natural sentences — no "...", no stage pauses. If drift happens anyway, regenerate (seed varies). Human ears verify every personal/shipping VO.
• First generation after app start loads Qwen 1.7B — slow on this Intel Mac; later ones are fine. Batch VO accordingly.
• Voicebox MCP is registered for the machine project (claude mcp → voicebox).

Timing law (Brent, 2026-07-23 — refined over Law #5 v1–v3)

• Don't cut the words to fix timing — the script's natural length is usually right (Brent: "it says exactly the right amount"). Fix ENTRY and PACING first.
• Start with a real gap: ~2000ms after video start (never <1500ms) — the hook lands, THEN the voice comes in.
• Hard ceiling only: narration must end ≥1.5s before the video ends — vo.sh enforces and prints a word budget ((video−delay−1.5)×2.3). That budget is the CEILING for trimming, not a target.
• Sync watch-out: if the voice describes something 10s after it left the screen, the fix is usually the composition's scene holds (or splitting the narration to land per-scene), not shorter text.
• Voice v1 of any new format gets Brent's ears before it ships.

Format rulings (Brent)

• AGENT LAWS ship with his voice. JOURNEY episodes stay text-films (style law) — narrated versions are a YouTube/compilation flavor, not the Shorts default.
• English VO = Voicebox. Thai-in-Brent's-voice = ElevenLabs multilingual (1.5M credits parked) if/when wanted — key must arrive via file in ~/Desktop/GitSync/ (the consolidated credentials file), never chat.
• Personal/family videos never get queued anywhere, voiced or not.