English · Français
Documentary filmmaker's AI toolkit: voice repair, VHS analysis, FCP workflows, OCR for handwritten documents.
Built for real productions, not demos.
Ismaël Joffroy Chandoutis, filmmaker and artist. César 2022, Cannes, IDFA, Hot Docs, Ars Electronica.
These tools were built during active production of feature documentaries. They solve problems I actually have.
Four independent tools, one repo. Each has its own README with full documentation.
cinema-ai-toolkit/
├── voice-pipeline/ # Voice repair for documentary subjects
├── vhs-pipeline/ # VHS/Hi8/miniDV analysis with a vision LLM → FCP markers
├── fcp-workflow/ # Final Cut Pro auteur workflow + agentic scripting integration
├── prison-writing/ # OCR + graphological analysis for handwritten documents
└── ETHICS.md # Ethics statement for documentary AI tools
Voice direction for documentary filmmakers working with non-professional subjects. The subject's voice is authentic and irreplaceable. This pipeline repairs the performance, not the voice.
RAW RECORDINGS → Denoise → Isolate → Diarize → Segment → Normalize
→ Enhance → Inpaint → Fix delivery → Voice clone ready
Stack: DeepFilterNet 3, Resemble Enhance, ElevenLabs, Sesame CSM, Chatterbox
Analyze hours of analog archive footage (VHS, Hi8, miniDV, Super8) with a vision LLM. Export colored markers directly to Final Cut Pro 12.
Archive footage → ffmpeg proxy → vision LLM analysis → FCPXML colored markers → FCP 12
| Marker Color | Meaning |
|---|---|
| Red | Strong interest, must review |
| Orange | Narrative structure moment |
| Blue | Standard marker |
| Green | Glitch / artifact |
Stack: Vision LLM (fast), ffmpeg, FCPXML
A-to-Z workflow for auteur cinema with Final Cut Pro, agentic scripting, and open source tools.
Philosophy: Hack everything. Local first. Plain text is king. Automate the boring parts.
OCR + graphological analysis + data mining from photographs of handwritten documents and prison correspondence. Built for documentary research.
For each image:
- Transcription: word-for-word OCR with illegibility markers
- Classification: letter, prison email, psychiatric report, legal document...
- Graphological analysis: pressure, slant, regularity, legibility
- Data mining: persons, locations, dates, themes, emotional state
Stack: fast vision LLM (primary) + high-tier vision LLM (recheck low-confidence)
| Model | Handwriting accuracy | Cost |
|---|---|---|
| Vision LLM (fast) | ~90% | $0.50/1M tokens |
| GPT-5 | ~90%+ | $$$ |
| Tesseract (local) | ~64% | Free |
All tools in this repo follow the ethics statement in ETHICS.md. Documentary AI tools operate on real people's lives. The technology is never neutral.
Each tool has its own requirements. See the README in each directory.
# Voice pipeline
cd voice-pipeline && pip install -r requirements.txt
# VHS pipeline
cd vhs-pipeline && pip install -r requirements.txtConsolidated from: cinema-voice-pipeline · vhs-ai-pipeline · fcp-auteur-workflow · prison-writing-analyzer