legends-captions

skill
Guvenlik Denetimi
Uyari
Health Uyari
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 5 GitHub stars
Code Gecti
  • Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Agentic video caption QA and subtitle rendering CLI: contextual ASR correction, word-level forced alignment, active-word ASS subtitles, and proof frames.

README.md

legends-captions

release
checks
license
Python
Agents

Video captions are easy to generate and painful to trust. Legends Captions is an agentic video caption system that treats transcript correction,
forced alignment, active-word timing, proof frames, and revision memory as one
release workflow instead of a pile of manual cleanup.

This is built for the gap between generic auto captions and real publishing
quality. The system is intentionally overbuilt: it uses context-aware correction
policy, Parakeet-quality ASR, NeMo Forced Aligner timing, ASS caption rendering,
manifest evidence, and final-render QA so the first pass is usually close and
the second pass teaches the next run.

What It Does

  • Builds accurate captions from the exact final video clock.
  • Applies contextual ASR correction before render.
  • Fixes model names, product names, acronyms, hardware names, and obvious phrase
    errors when the evidence is strong.
  • Keeps spoken acronyms such as L O R A independently highlightable when the
    speaker spells the term.
  • Uses forced alignment as final timing evidence when exact highlights matter.
  • Produces active-word ASS captions with consistent visual rules.
  • Preserves manifests so every text correction and timing fallback can be
    reviewed later.
  • Turns user fixes into reusable policy, config, and regression tests.

What Makes This Different

Generic auto captions Legends Captions
Trusts ASR text too much Treats ASR as evidence, not the editor
Often misses domain terms Uses glossary and contextual correction
Word timing can be rough Uses forced alignment when timing matters
Fixes disappear after the run Revisions become policy and tests
Proofing is manual and informal Requires manifests and proof-frame QA

The goal is not to make another one-click caption toy. The goal is a caption
deployment system that can take a real creator video, apply captions, catch the
obvious nonsense, and leave behind enough evidence for an agent to improve.

Quick Start

git clone https://github.com/avalonreset/legends-captions.git
Set-Location -LiteralPath .\legends-captions
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e .
.\.venv\Scripts\legends-captions.exe doctor

Try the policy engine:

.\.venv\Scripts\legends-captions.exe normalize --uppercase "low raw teaches the model"
.\.venv\Scripts\legends-captions.exe normalize --uppercase "it may not be as exacting 2.0,"
.\.venv\Scripts\legends-captions.exe normalize --manifest "L O R A."
.\.venv\Scripts\legends-captions.exe render-ass --uppercase "low raw teaches the model"
.\.venv\Scripts\legends-captions.exe ffmpeg-filter --ass "C:\work\captions.ass"

Run tests on any platform with Python 3.11+:

python -m unittest discover -s tests

Caption Policy Examples

Heard by ASR Display policy
low raw LoRA
C dance 2.0 Seedance 2.0
exacting 2.0 in AI-video context exciting as Seedance 2.0
what is this unlock what does this unlock
sticking the 30 sticking to 30
L O R A preserve per-letter highlight timing

Agent setup (via cto-legends)

Part of the CTO Legends ecosystem. cto-legends is the only registered skill; this repo vendors a pinned copy at skills/cto-legends/SKILL.md.

Install with cto-legends install legends-captions, then follow the module recipe the router loads. Do not register this module as its own skill.

Project Map

.raw/       preserved source notes, scripts, manifests, and run evidence
assets/     README banner and social preview
configs/    reusable caption style and correction defaults
docs/       install, architecture, compatibility, release, and QA docs
release/    Skool post drafts and promo copy (art staging stays local)
skills/     vendored `cto-legends` router skill copy (only registered skill)
src/        Python package for policy, correction, ASS helpers, and manifests
tests/      regression tests created from real caption fixes
wiki/       product brain for policies, playbooks, agents, and QA

Where Transcripts Come From

This module corrects, times, and renders captions. It does not record audio
or run ASR itself. Get transcripts from the Legends module that fits:

Then bring the text here for correction policy, forced-alignment timing,
active-word rendering, and proof evidence.

Documentation

Acknowledgments

The caption policy engine, CLI, manifests, docs, and release art
direction are original work by Avalon Reset, released under MIT. The
package has zero runtime dependencies.

Release art sources are AI-generated images composited locally with
scripts/stage_banner_concepts.py; no stock photography or third-party
artwork is shipped.

This module works alongside tools it does not ship: FFmpeg (LGPL/GPL),
faster-whisper (MIT), NVIDIA NeMo with Parakeet weights (Apache-2.0
code, CC-BY-4.0 weights), yt-dlp (Unlicense), and Montserrat (SIL Open
Font License, referenced by name in the default ASS style only).

FAQ

Does legends-captions bundle FFmpeg or ASR model weights?
No. It is a deterministic, zero-dependency caption policy, normalization, and .ass subtitle / manifest engine (doctor, policy-report, normalize, render-ass, ffmpeg-filter) that works alongside external tools (ffmpeg, faster-whisper, or NVIDIA NeMo Parakeet/NFA) on your machine without bundling model weights or spawning hidden shell wrappers.

Can I generate subtitles without burning them into a new video file?
Yes. legends-captions render-ass emits standalone styled .ass subtitle documents and JSON policy-correction manifests (--manifest-output), and legends-captions ffmpeg-filter --ass <path> formats a safely escaped libavfilter ass=filename='...' filtergraph string for when you choose to run an FFmpeg burn-in pass.

Status

Current release: v0.1.4.

License: MIT. See
LICENSE.

Repository target: avalonreset/legends-captions.


GitHub Star List: The complete cto-legends collection on one page. Bookmark or star the router and every module from a single GitHub list.

Specification & benchmarks on cto-legends.com:

Yorumlar (0)

Sonuc bulunamadi