Booth

Read your voiceover script out loud. A real-time AI director cuts your flubs, calls retakes with the exact line number, and prints the clean master when you wrap.

For anyone who records VO weekly (faceless channels, podcasters, ads), the take loop is 30-60 minutes of scrubbing per video: read, flub, hunt for where you were, fix it in an editor. Booth replaces the loop with a director on the other side of the glass.

Live: https://booth-voice.netlify.app · Event: AssemblyAI Voice Agent Hackathon (lablab.ai, Sep 2026)

---

Judge it in 90 seconds

1. Open the live link. On the landing, press "Run the bad take" — a bundled bad take runs through the real judge pipeline (sync transcription → line alignment → flub scoring) and stamps the retakes. No microphone needed.
2. Press Enter the booth →. A sample script is preloaded; or paste your own.
3. Take your position (mic permission), wait for the greeting.
4. Read three lines aloud. On line 3, deliberately say a wrong number (the script says "four cups" — say "two cups").
5. The director cuts in with the line number and correction; the slate stamps on the script.
6. Press That's a wrap → the master assembles (best read of every line) → Download master WAV. The flubbed read is gone; the retake is what printed.

Everything above is deterministic except steps 4-5, which are live voice.

What's real (and what isn't)

Claim	Real?	Evidence
Live voice agent listens, follows the script, speaks direction	Real	Managed Voice Agent API session; transcripts + reply audio streamed to the browser
Retake calls are the agent's own decisions	Real	call_retake / mark_take / finalize are client-side tools; every slate stamp on screen is a tool.call the agent made
Master assembled from word-level timestamps	Real	Sync API (universal-3.5-pro, timestamps: true); best-scoring read per line spliced with fades in-browser
Demo widget	Real pipeline	Runs sync transcription + alignment + scoring on bundled audio at press time; no mocks
Cut-ins fire mid-sentence	Honest gap	The agent replies after end-of-turn (you pause, then get cut) — a real director's beat, not literal barge-in. A client-side nudge (reply.create) forces the call when a flub ends in silence
The director cuts on every flub, every time	Honest gap	The director is an LLM: on rare reads it may skip a cut. Number flubs — the most common VO error — have a deterministic client-side backstop (numeric mismatch detection fires the cut 1.8 s after the pause)
Demo audio is a human	Honest gap	Synthesized on macOS for a deterministic demo; bring your own mic for the live path
"Print master" always perfect	Honest gap	Line selection is alignment-score based; if a line never gets a clean read it stays flubbed in the take log and the raw session WAV is provided

Architecture

Browser ── mic ── AudioWorklet ──► PCM16 24 kHz ──► Voice Agent WS (temp token, single-use)
   │                                   │                 │  reply.audio → speakers
   │                                   └─► full-rate master capture (Int16)
   │
   └─ wrap: master WAV ──► /api/sync ──► sync.assemblyai.com (universal-3.5-pro, timestamps)
                              words ──► align to lines ──► best-read splice ──► master WAV

• No persistent server. Two serverless routes (mint token, proxy sync). All audio processing is client-side Web Audio. Static hosting: Vercel.
• API key never reaches the browser in production; a BYO-key fallback exists for local runs.
• Session discipline: session.end → session.ended → close (no billable grace windows).

Sponsor primitives, and where they're load-bearing

Primitive	Used for	Removable?
Voice Agent API (inline config, client-side tools, interruption handling)	The director: listening, judging, speaking retake calls	No — Booth IS this agent
Sync STT, universal-3.5-pro, word timestamps	Master assembly: aligning reads to lines	No — the splice is made of these timestamps
Temporary session tokens	Browser-direct WS, zero server	No — the static-hosting shape depends on it

Run it

npm install
ASSEMBLYAI_API_KEY=<key> npm run dev   # http://localhost:3000
npm run build && npm start

Without a server key, the booth prompts for a BYO key (stored in localStorage, used only to mint tokens).

Stack

Next.js 16 · React 19 · Tailwind 4 · Web Audio (AudioWorklet capture, OfflineAudioContext splice, WAV encode) · AssemblyAI Voice Agent API + Sync API. No audio libraries.

Credits

Built by Raphie for the AssemblyAI Voice Agent Hackathon. Palette borrowed from AssemblyAI's own verified brand tokens.