AI video creator builds voice recalibration pipeline to fix model voice-swapping bugs
letandrewcook · x · 2026-10-10
Andrew released a new episode of his AI-produced series The One With the Constrained Allocation Problem, the first with a credit roll, and says anything over two minutes gets credits from now on.
The hard part was audio: when two characters share a scene, the AI video model gets confused and swaps their voices, and invents a voice when it doesn't recognize someone. He had to build a voice recalibration pipeline — tricky because the original and AI-generated voices differ slightly, so recalibrating with another voice model always failed. A fallback was 480p video-to-video generation with single-actor dialogue, though that's expensive. A few intentional mistakes were left in.
Related event: Indie Creator Builds Voice Recalibration Pipeline for AI-Made Sitcom(2 posts)→
More from coding & agent
- Cerebral Valley and Crusoe host Recursive Agents Hackathon with $25K in prizes — PolarBearby · 2026-10-10
- PM with Claude Code finds potential exoplanet after 74 adversarial analyses — sujingshen · 2026-10-10
- Open-source agent Comma runs 24/7, handling email, Linear and Notion tasks for you — HeyAmit_ · 2026-10-10
- Andrew Ng releases 1-hour course on building agentic knowledge graphs from scratch — nikola_mr64990 · 2026-10-10
- Mnemos: an engram-inspired memory architecture doubles short-term recall vs Claude's native memory — RileyRalmuto · 2026-10-10
- LLM reliability #6-7: guardrails, fallbacks, and production monitoring — goyalshaliniuk · 2026-10-10