Slides To Video

Slides to Video

Turn already-prepared slides into a narrated, captioned MP4 in a user-chosen language and presenter role. Everything is free of charge and runs from a dedicated Python virtualenv plus ffmpeg. Voice quality is maximized with a language-aware TTS ladder; English and Vietnamese are first-class.

Windows Runtime Commands

On native Windows, use the managed Windows runner and the native runtime command target. Set $runtime to the installed runtime root. Multi-agent installs usually use %LOCALAPPDATA%\ai-agents-skills\runtime.

$runtime = if ($env:AAS_RUNTIME_ROOT) { $env:AAS_RUNTIME_ROOT } else { "$env:LOCALAPPDATA\ai-agents-skills\runtime" }
& "$runtime\run_skill.ps1" "skills/slides-to-video/run_slides_to_video.ps1" <args>
& "$runtime\run_skill.ps1" "skills/slides-to-video/run_slides_to_video.ps1" <args>

POSIX examples below use run_skill.sh and the .sh command target.

When to use

Use this skill when the user wants to:

• narrate a deck they already exported as high-resolution PNG (1920x1080 or 3840x2160), PDF, or PPTX
• produce a presentation/lecture video with a chosen voice, language, and role (presenter, teacher, lecturer, narrator, pitch)
• add Tier-1 visual effects (Ken Burns, highlight, spotlight, laser pointer, step reveal) on top of the slides
• generate synced captions (SRT + VTT), with Vietnamese diacritics handled

Do NOT use this for authoring math animations from scratch (handwritten equations, equation morphing) -- that is the planned, separate slides-to-video Manim extension. This skill consumes pre-made slides.

Free-tools and licensing posture

• TTS: edge-tts (online, best voice), Kokoro (offline, Apache-2.0), Piper (offline, 30+ languages). XTTS-v2 is excluded (non-commercial license).
• Assembly: ffmpeg is the one required system tool. Use an LGPL build; the H.264 patent obligation applies to redistribution of output, not local use.
• Python packages run only in the dedicated venv; nothing is vendored.

Setup (one time, explicit)

Step 1 - install the SYSTEM tools with your OS package manager. ffmpeg/ffprobe is the only hard requirement; espeak-ng is needed only for offline TTS (Kokoro/Piper); a PPTX renderer is needed only for PPTX input. On Windows, Microsoft PowerPoint from Microsoft Office is supported through COM automation, so LibreOffice is not needed when Office is installed. On Linux/macOS, use LibreOffice for PPTX input.

# Debian/Ubuntu
sudo apt-get update && sudo apt-get install -y ffmpeg espeak-ng libreoffice
# Fedora
sudo dnf install -y ffmpeg espeak-ng libreoffice
# Arch
sudo pacman -S ffmpeg espeak-ng libreoffice
# macOS (Homebrew)
brew install ffmpeg espeak-ng && brew install --cask libreoffice
# Windows (winget); Microsoft Office satisfies PPTX input if installed
winget install Gyan.FFmpeg eSpeak-NG.eSpeak-NG

Step 2 - create the dedicated venv and check readiness. Once the skill is installed into an agent home, use the managed runner:

bash "${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh" skills/slides-to-video/run_slides_to_video.sh setup
bash "${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh" skills/slides-to-video/run_slides_to_video.sh doctor

To try it straight from the repo before installing, call the script directly:

bash canonical/runtime/skills/slides-to-video/run_slides_to_video.sh setup
bash canonical/runtime/skills/slides-to-video/run_slides_to_video.sh doctor

setup creates ~/.local/share/slides-to-video-venv and installs requirements. doctor prints a JSON report of exactly what is present or missing and never installs anything. On Windows, doctor reports system_tools.powerpoint when Microsoft PowerPoint is available for PPTX rendering.

The three-phase, human-in-the-loop flow

Rendering is impossible until the transcript is explicitly approved; editing the transcript after approval re-blocks rendering (the approval is pinned to the SHA-256 of transcript.json).

Phase 1 - Analyze. Ingest slides and record decisions.

... run_slides_to_video.sh analyze \
  --input "/path/to/slides_dir_or_file" \
  --work-dir "/path/to/deck_work" \
  --language vi-VN --role teacher --engine-policy auto --resolution 1920x1080

Writes deck.json, config.json, and ordered frames under slides/. For PNG-only input there is no embedded text, so seed each slide's narration by reading the slide image (vision) before/while drafting.

Phase 2 - Draft transcript and iterate. Scaffold, then write the narration.

... run_slides_to_video.sh draft --work-dir "/path/to/deck_work"

This creates transcript.json (authoritative spoken text + per-slide effects) and transcript.md (human-readable view). Edit transcript.json to write the actual narration per slide, then regenerate the view and present it to the user. For math decks, run verbalize to rewrite notation into speakable words (e.g. x^2 -> "x squared" / "x binh phuong"); the user reviews the spoken form. Iterate (edit specific slides, regenerate) until the user is satisfied.

Approve (explicit gate).

... run_slides_to_video.sh approve --work-dir "/path/to/deck_work"

Only do this on an explicit user "approve/render it". It pins the transcript SHA.

Phase 3 - Render.

... run_slides_to_video.sh render --work-dir "/path/to/deck_work"

Per slide: synthesize narration via the language-aware ladder, normalize to WAV, measure the exact duration with ffprobe, build a clip of exactly that length with any effects, then losslessly concatenate. Outputs video.mp4, captions.srt/.vtt, and render_report.json. Refuses if not approved.

status --work-dir ... reports pipeline state at any time.

Languages

The core is language-agnostic. English and Vietnamese are tuned (math lexicon, voices, fonts). Caption font selection prefers the lexicon's font hint, then falls back to installed Vietnamese-capable fonts such as Be Vietnam Pro, Arial, Calibri, Segoe UI, Tahoma, Times New Roman, or DejaVu Sans. Any other language the engines cover is supported generically: edge-tts serves ~70 locales online; Piper covers 30+ offline; Kokoro covers 9. The ladder is language-aware -- e.g. Vietnamese never routes into Kokoro (which has no Vietnamese voice); offline Vietnamese uses Piper vi_VN-vais1000-medium, online uses edge-tts vi-VN-HoaiMyNeural/NamMinhNeural. To make another language first-class, add s2v/languages/<lang>.json (voices + math lexicon) and confirm a covering font; no code change.

Effects

Tier-1 effects operate on the rendered slide pixels (no authoring): ken_burns, highlight, spotlight, laser, reveal. Specify them per slide in transcript.json under each slide's effects list. See references/effects.md.

Animated math interludes (Manim)

A slide's visual can be a pre-rendered video clip instead of a static image: set clip_path on a slide in transcript.json, or insert one with add-interlude. This is how a manim-math-animation clip is mixed into a narrated deck.

# 1) render a silent Manim clip with the manim-math-animation skill, then:
... run_slides_to_video.sh add-interlude --work-dir "/path/to/deck_work" \
  --clip "/path/to/manim_clip.mp4" --after 0 \
  --transcript "Completing the square gives a perfect square."

The clip becomes a new slide (inserted after the given index; slides are reindexed and the deck/transcript stay 1:1). Narration is still synthesized by the TTS ladder, so the Manim clip stays silent (one timing owner). The segment runs to max(clip_duration, narration_duration): a shorter clip freezes its last frame, shorter narration is padded with silence -- so the lossless concat stays drift-free. add-interlude invalidates any prior approval, so re-approve before render. See references/pipeline-and-sync.md.

References

• references/pipeline-and-sync.md - duration-driven sync, the 1:1 pairing contract, encoding settings, and the approval gate.
• references/effects.md - effect types and their parameters.
• references/languages-and-tts.md - the engine ladder, per-language data, and spoken-math verbalization.
• references/design.md - architecture, scope, and the deferred Manim extension.