accessibility-pack
Accessibility pack for any finished Video Studio edit: SDH subtitles with [sound] tags and speaker IDs, broadcast-rule caption files (SRT, VTT, TTML), a clean transcript, an audio-description (AD) script written from the picture, and an AD voice track mixed into a separate descri
Install
npx skills add https://github.com/Viprasol-Tech/claude-video-editing-skills/tree/main/plugins/video-studio/skills/accessibility-pack
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install viprasol-tech-claude-video-editing-skills@llmmart
git clone https://github.com/Viprasol-Tech/claude-video-editing-skills.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole viprasol-tech/claude-video-editing-skills collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
Accessibility Pack
The same film for people who can't hear it and people who can't see it, delivered as files a broadcaster or LMS will accept.
- Works on: any finished edit with a plan (
W/plan.json+W/final.mp4), or any video plus a words file - Makes:
captions.srt/.vtt/.ttml(SDH),transcript.txt,ad_script.txt,ad_*.wav,final_ad.mp4(described version), QC reports - Standards used: Netflix/BBC caption practice (42 chars x 2 lines, <= 20 cps adults / 17 kids, 0.833-7 s), WCAG 2.2 (1.2.2 captions, 1.2.3/1.2.5 audio description, 1.2.8 media alternative = transcript)
Before you start
- Needs video-studio-core in the same skills folder (
CORE). First run:python CORE/scripts/check_env.py. - Start from a locked edit: captions and AD are timed to picture; any re-cut means redoing them.
- Ask: who speaks (names for speaker IDs), the language, whether the target is web (WCAG) or a broadcaster (which spec), whether a described version is wanted, and who will review the AD script (ideally a blind or low-vision reviewer).
- Honest limits: no SCC/CEA-608 encoder (say so if a broadcaster asks for SCC), no automatic speaker diarisation, no automatic sound-event detection, and finish.py does not duck the programme under AD lines. The steps below work around each.
SDH captions (what makes them SDH, not plain subtitles)
- Speaker IDs when the speaker is off screen or not obvious:
PRIYA: ...(or[Priya]), at the start of the cue. - Sound tags in square brackets for sound that carries meaning:
[applause],[upbeat music],[phone buzzes],[laughs],[silence]. Describe the sound, not your opinion of it; skip sounds that mean nothing. - Music:
[upbeat music]at the start; lyrics in the cue with ♪ if they matter to the story. - Exact words, including meaningful hesitations; correct names and numbers; same cue length rules as broadcast.
Make them: the engine builds cues from a words file, so add the tags and IDs as extra "words" with times:
python - <<'EOF'
import json
p = json.load(open("W/plan.json", encoding="utf-8"))
w = [dict(x) for x in p["words"]]
w[0]["w"] = "PRIYA: " + w[0]["w"] # speaker ID on her first cue
sdh = [{"w": "[upbeat music]", "start": 0.0, "end": 0.9}] + w
d = p["output"]["duration"]
sdh.append({"w": "[applause]", "start": d - 0.9, "end": d - 0.1})
json.dump({"language": "en", "words": sdh}, open("W/sdh_words.json", "w", encoding="utf-8"), indent=1)
EOF
python CORE/scripts/deliver.py W/final.mp4 --to youtube --words W/sdh_words.json --out-dir W/access --burn no --no-thumbs --no-post
Give the video plus --words: with a plan as input, deliver.py uses the plan's own words and ignores --words.
A tag less than 0.7 s before speech joins the first spoken cue ([upbeat music] PRIYA: Here are three); for a cue of its
own, leave a gap of 0.8 s or more, or edit W/access/captions.srt by hand afterwards (it is plain text: keep the
numbering and timings, change the words). Children's content: add --kids (17 cps).
Transcript (WCAG 1.2.8): captions.srt text without timings, with speaker names and sound tags, plus a short
description of important visuals. Write it to W/access/transcript.txt.
Burned-in (open) captions for social feeds: --burn yes on the social presets; keep sidecars for YouTube/LMS.
Audio description
- Find the gaps. Read
W/plan.jsonwords: pauses of >= 1.5 s are AD slots. List what the picture shows that the sound doesn't say: on-screen text and slides, who appears, actions, settings, charts, the end card. - Write
ad_script.txtin present tense, objective, short ("A slide lists three numbered mistakes."). Never interpret emotion the face doesn't clearly show; name people once they are introduced; read on-screen text aloud. One line per slot, with its start time:14.4 | A slide lists three numbered mistakes. - Voice each line (offline TTS; Windows SAPI is always present, kokoro is better if installed):
python CORE/scripts/tts.py --text "A slide lists three numbered mistakes." --out W/ad/ad_01.wav --backend sapi --rate 1.1
python CORE/scripts/tts.py --list-voices
Check each WAV's length (ffprobe) fits its slot. Too long: shorten the words first, then --rate up to 1.2.
4. Mix into a described copy of the plan (same folder, so relative paths keep working):
python - <<'EOF'
import json
p = json.load(open("W/plan.json", encoding="utf-8"))
p["sfx"] = p.get("sfx", []) + [{"t": 14.4, "file": "ad/ad_01.wav", "gain": 1.0}]
p["outputFile"] = "final_ad.mp4"
json.dump(p, open("W/plan_ad.json", "w", encoding="utf-8"), indent=2)
EOF
python CORE/scripts/finish.py W/plan_ad.json --platform youtube
finish.py re-uses W/composite.mp4 (picture unchanged) and writes W/final_ad.mp4. The AD line plays on top of the
programme without ducking it, so place lines only in real pauses. No pause long enough? Make one: add
"freeze": [{"at": 14.4, "dur": 3.0}] (extended AD) to that clip in a timeline spec and rebuild, or shorten the line.
5. Optional AD-only track for players that take a separate AD audio stream: mix the AD WAVs onto silence with FFmpeg
(adelay per line + amix) and deliver it next to final_ad.mp4.
Broadcast caption rules (checked automatically)
deliver.py builds cues to Netflix/BBC practice and qc.py checks them: max 42 characters per line and 2 lines, <= 20 cps
(17 with --kids), 0.833-7 s per cue, 2-frame gaps, line breaks not after articles or between first and last name.
TTML is IMSC1 text profile. SCC (CEA-608) is not produced.
Check before delivering
python CORE/scripts/qc.py W/access/final_youtube.mp4 --plan W/plan.json --platform youtube --out-dir W/access/qc
python CORE/scripts/qc.py W/final_ad.mp4 --plan W/plan_ad.json --platform youtube --out-dir W/access/qc_ad
python CORE/scripts/audio_chain.py measure W/final_ad.mp4 --platform youtube
- Read
captions.srtstart to end against the audio: names, numbers, speaker IDs, sound tags in brackets. - Every AD line sits in a pause, never over dialogue; listen to the described version with your eyes closed.
- The described version's loudness passes (the AD lines raise the level; finish.py re-measures the whole mix).
- On-screen text that matters is either spoken by the presenter or read by AD.
- Deliver: SDH
captions.srt/.vtt/.ttml,transcript.txt,ad_script.txt,final_ad.mp4, and the QC reports; say what you could not do (SCC, programme ducking, a human AD review).
Example requests
- "Make SDH captions for my keynote, mark the applause and music, I'm Priya."
- "We need an audio-described version of this product video for our accessibility policy."
- "Give me caption files and a transcript the LMS will accept."
Files (claude-video-editing-skills)
-
SKILL.md 7.4 KB
--- name: accessibility-pack description: "Accessibility pack for any finished Video Studio edit: SDH subtitles with [sound] tags and speaker IDs, broadcast-rule caption files (SRT, VTT, TTML), a clean transcript, an audio-description (AD) script written from the picture, and an AD voice track mixed into a separate described version. Use when the user says 'make it accessible', 'add SDH captions', 'subtitles for the deaf and hard of hearing', 'closed captions file', 'audio description', 'described version', 'WCAG captions', 'caption file for LMS / broadcaster', 'transcript for the video'. Runs on video-studio-core (deliver.py captions, tts.py, finish.py mix, qc.py). Works after any genre skill." license: MIT --- # Accessibility Pack The same film for people who can't hear it and people who can't see it, delivered as files a broadcaster or LMS will accept. - **Works on:** any finished edit with a plan (`W/plan.json` + `W/final.mp4`), or any video plus a words file - **Makes:** `captions.srt/.vtt/.ttml` (SDH), `transcript.txt`, `ad_script.txt`, `ad_*.wav`, `final_ad.mp4` (described version), QC reports - **Standards used:** Netflix/BBC caption practice (42 chars x 2 lines, <= 20 cps adults / 17 kids, 0.833-7 s), WCAG 2.2 (1.2.2 captions, 1.2.3/1.2.5 audio description, 1.2.8 media alternative = transcript) ## Before you start - Needs **video-studio-core** in the same skills folder (`CORE`). First run: `python CORE/scripts/check_env.py`. - Start from a **locked** edit: captions and AD are timed to picture; any re-cut means redoing them. - Ask: who speaks (names for speaker IDs), the language, whether the target is web (WCAG) or a broadcaster (which spec), whether a described version is wanted, and who will review the AD script (ideally a blind or low-vision reviewer). - Honest limits: no SCC/CEA-608 encoder (say so if a broadcaster asks for SCC), no automatic speaker diarisation, no automatic sound-event detection, and finish.py does not duck the programme under AD lines. The steps below work around each. ## SDH captions (what makes them SDH, not plain subtitles) - **Speaker IDs** when the speaker is off screen or not obvious: `PRIYA: ...` (or `[Priya]`), at the start of the cue. - **Sound tags** in square brackets for sound that carries meaning: `[applause]`, `[upbeat music]`, `[phone buzzes]`, `[laughs]`, `[silence]`. Describe the sound, not your opinion of it; skip sounds that mean nothing. - **Music:** `[upbeat music]` at the start; lyrics in the cue with ♪ if they matter to the story. - Exact words, including meaningful hesitations; correct names and numbers; same cue length rules as broadcast. **Make them:** the engine builds cues from a words file, so add the tags and IDs as extra "words" with times: ```bash python - <<'EOF' import json p = json.load(open("W/plan.json", encoding="utf-8")) w = [dict(x) for x in p["words"]] w[0]["w"] = "PRIYA: " + w[0]["w"] # speaker ID on her first cue sdh = [{"w": "[upbeat music]", "start": 0.0, "end": 0.9}] + w d = p["output"]["duration"] sdh.append({"w": "[applause]", "start": d - 0.9, "end": d - 0.1}) json.dump({"language": "en", "words": sdh}, open("W/sdh_words.json", "w", encoding="utf-8"), indent=1) EOF python CORE/scripts/deliver.py W/final.mp4 --to youtube --words W/sdh_words.json --out-dir W/access --burn no --no-thumbs --no-post ``` Give the **video** plus `--words`: with a plan as input, deliver.py uses the plan's own `words` and ignores `--words`. A tag less than 0.7 s before speech joins the first spoken cue (`[upbeat music] PRIYA: Here are three`); for a cue of its own, leave a gap of 0.8 s or more, or edit `W/access/captions.srt` by hand afterwards (it is plain text: keep the numbering and timings, change the words). Children's content: add `--kids` (17 cps). **Transcript (WCAG 1.2.8):** `captions.srt` text without timings, with speaker names and sound tags, plus a short description of important visuals. Write it to `W/access/transcript.txt`. **Burned-in (open) captions** for social feeds: `--burn yes` on the social presets; keep sidecars for YouTube/LMS. ## Audio description 1. **Find the gaps.** Read `W/plan.json` `words`: pauses of >= 1.5 s are AD slots. List what the picture shows that the sound doesn't say: on-screen text and slides, who appears, actions, settings, charts, the end card. 2. **Write `ad_script.txt`** in present tense, objective, short ("A slide lists three numbered mistakes."). Never interpret emotion the face doesn't clearly show; name people once they are introduced; read on-screen text aloud. One line per slot, with its start time: `14.4 | A slide lists three numbered mistakes.` 3. **Voice each line** (offline TTS; Windows SAPI is always present, kokoro is better if installed): ```bash python CORE/scripts/tts.py --text "A slide lists three numbered mistakes." --out W/ad/ad_01.wav --backend sapi --rate 1.1 python CORE/scripts/tts.py --list-voices ``` Check each WAV's length (`ffprobe`) fits its slot. Too long: shorten the words first, then `--rate` up to 1.2. 4. **Mix into a described copy** of the plan (same folder, so relative paths keep working): ```bash python - <<'EOF' import json p = json.load(open("W/plan.json", encoding="utf-8")) p["sfx"] = p.get("sfx", []) + [{"t": 14.4, "file": "ad/ad_01.wav", "gain": 1.0}] p["outputFile"] = "final_ad.mp4" json.dump(p, open("W/plan_ad.json", "w", encoding="utf-8"), indent=2) EOF python CORE/scripts/finish.py W/plan_ad.json --platform youtube ``` `finish.py` re-uses `W/composite.mp4` (picture unchanged) and writes `W/final_ad.mp4`. The AD line plays on top of the programme without ducking it, so place lines only in real pauses. No pause long enough? Make one: add `"freeze": [{"at": 14.4, "dur": 3.0}]` (extended AD) to that clip in a timeline spec and rebuild, or shorten the line. 5. **Optional AD-only track** for players that take a separate AD audio stream: mix the AD WAVs onto silence with FFmpeg (`adelay` per line + `amix`) and deliver it next to `final_ad.mp4`. ## Broadcast caption rules (checked automatically) `deliver.py` builds cues to Netflix/BBC practice and `qc.py` checks them: max 42 characters per line and 2 lines, <= 20 cps (17 with `--kids`), 0.833-7 s per cue, 2-frame gaps, line breaks not after articles or between first and last name. TTML is IMSC1 text profile. SCC (CEA-608) is not produced. ## Check before delivering ```bash python CORE/scripts/qc.py W/access/final_youtube.mp4 --plan W/plan.json --platform youtube --out-dir W/access/qc python CORE/scripts/qc.py W/final_ad.mp4 --plan W/plan_ad.json --platform youtube --out-dir W/access/qc_ad python CORE/scripts/audio_chain.py measure W/final_ad.mp4 --platform youtube ``` - Read `captions.srt` start to end against the audio: names, numbers, speaker IDs, sound tags in brackets. - Every AD line sits in a pause, never over dialogue; listen to the described version with your eyes closed. - The described version's loudness passes (the AD lines raise the level; finish.py re-measures the whole mix). - On-screen text that matters is either spoken by the presenter or read by AD. - Deliver: SDH `captions.srt/.vtt/.ttml`, `transcript.txt`, `ad_script.txt`, `final_ad.mp4`, and the QC reports; say what you could not do (SCC, programme ducking, a human AD review). ## Example requests - "Make SDH captions for my keynote, mark the applause and music, I'm Priya." - "We need an audio-described version of this product video for our accessibility policy." - "Give me caption files and a transcript the LMS will accept."
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.