Video should be able to explain itself.
Semantic Video is an open ecosystem for carrying transcripts, timing, visible text, scenes, relationships, embeddings, provenance, and validation alongside the media they describe.
The goal is simple: give video tools, editors, search systems, automation pipelines, and AI agents a shared observation layer so they can understand media without repeatedly decoding and re-analyzing it from scratch.
SVP Core stores observations, not interpretations. AI may consume SVP, but AI does not define SVP.
| Project | What it is |
|---|---|
| SVP | The open specification and reference toolchain for building, validating, inspecting, and querying semantic video packages. |
| SVP for macOS | Native file support, Quick Look previews, Finder actions, transcript viewing, package inspection, and system media access. |
- SVP — source media and semantic observations together in one portable package.
- SVPI — a media-bound semantic sidecar that keeps the source media where it already lives.
- Embedded SVPI — one canonical SVPI carried inside a supported MP4, MOV, M4V, or M4A file.
An SVP package can carry source media, transcripts, word timing, speakers, visible text, colors, shots, scenes, entities, relationships, depth, embeddings, search indexes, provenance, and validation records.
The format is strict, machine-readable, and model-agnostic. Its contents are designed to remain inspectable and its claims are designed to remain testable.
There is a public roadmap. SVP is being built in visible, validator-backed stages, and the repository documents both the foundation already in place and the work still required to harden the 1.0 standard and toolchain.
- Specification, schemas, and registries
- Package builder, validator, inspector, and query tools
- SVPI sidecars and embedded SVPI transport
- Model-bundle verification and installation
- A reproducible Apple Silicon macOS build path
- Deterministic entity identity and track quality
- Relationship traversal and graph-health hardening
- Mask, spatial-region, scene, and timeline foundations
- Transcript, OCR, color, vision, and embedding quality
- Stricter validator coverage and package hygiene
- Entity performance profiles and inference controls
- Library Search and a Librarian daemon
- More efficient in-package SQLite queries
Follow the implementation roadmap, the active implementation focus, and open issues on GitHub.
Start with the SVP repository to read the standard, build the reference tools, create a package, run the validator, or inspect the format in detail.
Semantic Video is open source. The reference implementation is licensed under Apache 2.0, and the specifications, schemas, and registries are dedicated under CC0 1.0 Universal.