A regional sports network wants to broadcast a live basketball game. Today, that requires a production truck, a technical director, camera operators, audio engineers, and a graphics team. The cost: $50,000-$200,000 per event. With V100's AI Director, it requires cameras and an internet connection.
The 20Hz Decision Loop
V100's AI Director runs a scoring loop 20 times per second across all camera feeds. Each tick:
- Face detection via ONNX Runtime (GPU-accelerated) identifies how many faces are visible and how prominent the main subject is in each camera
- Audio energy analysis via FFT determines which mic is picking up the most relevant sound (speaker, crowd reaction, referee whistle)
- Motion analysis detects fast action moments that demand wider shots or quick cuts
- Composite scoring weights all signals:
0.35*audio + 0.30*face + 0.20*motion + 0.15*speaking
Natural Pacing, Not Robotic Switching
The difference between AI Director and simple "loudest speaker" switching is the pacing model. Real broadcast directors don't cut every time something happens — they follow a rhythm. V100's pacing engine uses a logistic urgency curve that builds over time, respects minimum shot durations (no flickering), enforces maximum shot durations (no dead air), and avoids cutting mid-sentence.
Combined with Voice Dubbing: 40 Languages, Zero Crew
The real disruption happens when AI Director combines with V100's voice dubbing engine. A single live event can be simultaneously broadcast in 40+ languages — each with the original commentator's voice cloned via XTTS-v2 neural TTS. The pipeline runs at under 2 seconds end-to-end: Deepgram STT (100ms) + translation (200ms) + TTS synthesis (500ms) + audio mix with ducking.
For regional sports rights, this is an entirely new business model. Instead of licensing exclusive rights per territory, a network can license globally and let V100 handle the localization automatically.
Production Crew Override
AI Director doesn't replace human judgment — it augments it. At any point, a production director can take manual control via the override API. The AI pauses until released, then seamlessly resumes autonomous direction. This makes V100 suitable for both fully automated broadcasts (small events, practice sessions) and hybrid production (major events where a human makes final calls on replays and graphics).
The API
8 endpoints handle the full lifecycle:
POST /api/v1/director/sessions — Create session with cameras + rules
GET /api/v1/director/sessions/{id} — Live camera scores + pacing metrics
POST /api/v1/director/sessions/{id}/rules — Update rules mid-broadcast
POST /api/v1/director/sessions/{id}/override — Manual takeover
POST /api/v1/director/sessions/{id}/release — Resume AI
GET /api/v1/director/sessions/{id}/history — Cut decision log
POST /api/v1/director/sessions/{id}/stop — End session
GET /api/v1/director/stats — Service metricsWhat This Means for Broadcast
No platform — Twilio, AWS IVS, Mux, Agora — offers automated multi-camera direction at the relay layer. V100 is the only broadcast infrastructure where the AI that switches cameras also controls the PPV access tokens, geo-gating, blackout enforcement, wagering metadata, and ad insertion. It's a single platform replacing an entire production stack.