slidesonnet 0.2.0

Compile text-based presentations into narrated videos

Compile text-based slide presentations into narrated MP4 videos.

Write your slides in MARP Markdown or LaTeX Beamer, add narration with <!-- say: --> comments, and slideSonnet handles TTS synthesis, video composition, and assembly — with incremental builds that only re-synthesize changed slides.

How it works

slidesonnet.yaml (playlist)
    |
    ├── 01-intro/slides.md   → [parse → TTS → compose] → module_01.mp4
    ├── animations/euler.mp4  → [passthrough]            → module_02.mp4
    ├── 02-proofs/slides.tex  → [parse → TTS → compose] → module_03.mp4
    └── [assemble] ─────────────────────────────────────→ my-course.mp4

A playlist file chains modules together — MARP slides, Beamer slides, and pre-existing video files. Each module is built independently, then concatenated into the final video. pydoit manages the build graph with content-hash caching, so only changed slides trigger TTS.

Installation

External dependencies

Install these system packages first:

Tool	Required?	What it does	Install
ffmpeg	Yes	Video composition and concatenation	sudo apt install ffmpeg
marp-cli	Yes (for MARP slides)	Converts Markdown slides to PNG images	npm install -g @marp-team/marp-cli
latexmk + pdflatex + pdftoppm	Only for Beamer	Compiles LaTeX (latexmk drives pdflatex to convergence) and extracts slide images	sudo apt install latexmk texlive-latex-base poppler-utils

After installing, run slidesonnet doctor to verify everything is set up correctly.

Install slideSonnet

With uv (recommended):

uv tool install slidesonnet[piper]

With pipx:

pipx install slidesonnet[piper]

The [piper] extra includes Piper TTS for free local speech synthesis. Omit it if you plan to use ElevenLabs instead.

Quick start

# Create an example project (MARP Markdown)
slidesonnet init md myproject

# Build the video
cd myproject
slidesonnet build

Example: The Basel Problem

A 10-minute narrated lecture on the Basel Problem, built entirely from a single slides.md file:

Source: examples/basel-problem/

Showcase example

The examples/showcase/ directory is a single-file MARP presentation introducing slideSonnet through a dialog between two voices. It demonstrates narration, fragment animation, voice switching, silent/skipped slides, math, code, and images — all in one slides.md file.

Source: examples/showcase/ — includes pronunciation dictionaries and a playlist with ElevenLabs and Piper voice configuration.

Writing slides

MARP Markdown

Add narration with <!-- say: --> HTML comments:

---
marp: true
---

# Introduction

<!-- say: Welcome to the lecture. Today we cover graph theory basics. -->

---

# Euler's Theorem

<!-- say(voice=alice): Let me explain this theorem carefully. -->

---

# Diagram

<!-- nonarration -->

---

# Hidden Notes

<!-- skip -->

Annotation	Effect
<!-- say: text -->	Narrate with default voice
<!-- say(voice=alice): text -->	Narrate with a named voice preset
<!-- nonarration -->	Show slide with silence (uses global silence_duration)
<!-- nonarration(5) -->	Show slide with silence for 5 seconds (per-slide override)
<!-- skip -->	Omit slide from video entirely
(none)	Treated as silent, emits a warning

Multi-line narration is supported. Slides with multiple <!-- say: --> directives are expanded into animated sub-slides with progressive fragment reveal — see MARP documentation for details.

Beamer LaTeX

Use the \say command (defined as a no-op by slidesonnet.sty so LaTeX compiles normally). The <N> mirrors beamer's own overlay specs — it picks which built-up step of the frame the narration plays on:

\usepackage{slidesonnet}

\begin{frame}
  \frametitle{Euler's Theorem}
  \say<1>{The sum of all vertex degrees equals twice the number of edges.}
  \onslide<2->{\say<2>[voice=alice]{Now watch what happens as the graph grows.}}
\end{frame}

Beamer forms: \say<N>{}, \say<N>[voice=alice]{}, \nonarration, \nonarration[5] (per-slide duration override), \slidesonnetskip. Every \say must carry a step number — <N> (recommended) or the legacy [N] / [slide=N] bracket form; a bare \say{} is rejected. Each frame is split into one video segment per beamer overlay step (counted from the compiled .nav, so \pause, \onslide<>, \item<> all work); steps without a \say are held silently — see Beamer documentation for details.

Playlist format

A single .yaml file per presentation. Configuration and module list in pure YAML:

title: Graph Theory Lecture 1
tts:
  backend: piper
  piper:
    model: en_US-lessac-medium
  elevenlabs:
    api_key_env: ELEVENLABS_API_KEY
    voice_id: pNInz6obpgDQGcFmaJgB
voices:
  alice:
    piper: en_US-amy-medium
    elevenlabs: 21m00Tcm4TlvDq8ikWAM
pronunciation:
  shared:
    - pronunciation/cs-terms.md
    - pronunciation/math-terms.md
  # piper:
  #   - pronunciation/piper-hacks.md
  # elevenlabs:
  #   - pronunciation/elevenlabs-hacks.md
video:
  resolution: 1920x1080
  fps: 24
  crf: 23
  pad_seconds: 1.5
  pre_silence: 1.0
  silence_duration: 3.0
  crossfade: 0.5
modules:
  - 01-intro/slides.md
  - animations/euler.mp4
  - 02-proofs/slides.tex
  - 03-summary/slides.md

• Module type is auto-detected from extension (.md → MARP, .tex → Beamer, .mp4 / .mkv / .webm / .mov → video passthrough)
• Lines starting with // are comments (filtered before YAML parsing)
• Video files are used as-is

Pronunciation files

Reusable .md files with **word**: replacement pairs:

# CS Pronunciation Guide

## People
**Dijkstra**: DYKE-struh
**Euler**: OY-ler

## Terms
**adjacency**: uh-JAY-suhn-see

Replacements are word-boundary aware (won't change "Eulerian") and case-insensitive. Reference them in the playlist under pronunciation:.

Per-backend pronunciation

Pronunciation workarounds that fix one TTS engine often break another. You can specify separate files per backend:

pronunciation:
  shared:
    - pronunciation/names.md
  piper:
    - pronunciation/piper-hacks.md
  elevenlabs:
    - pronunciation/elevenlabs-hacks.md

When building with --tts piper, the effective dictionary is shared + piper. With --tts elevenlabs, it's shared + elevenlabs. Backend-specific entries override shared entries for the same word.

The flat list format still works and is treated as shared:

pronunciation:
  - pronunciation/names.md

Voice presets

Define named voices in the playlist. Each preset can map to different voice IDs per TTS backend, so --tts piper and --tts elevenlabs both resolve correctly:

voices:
  alice:
    piper: en_US-amy-medium
    elevenlabs: 21m00Tcm4TlvDq8ikWAM
  bob:
    piper: en_US-joe-medium
    elevenlabs: pNInz6obpgDQGcFmaJgB

A simple string value is also supported — it is used as-is regardless of backend:

voices:
  alice: en_US-amy-medium

Then use presets per-slide: <!-- say(voice=alice): ... -->. If a preset has no mapping for the active backend, the slide falls back to the default voice with a warning.

API keys

For ElevenLabs, store keys in a .env file at the project root (auto-loaded at build time):

ELEVENLABS_API_KEY=sk-xxx-your-key

The playlist references env var names, never values: api_key_env: ELEVENLABS_API_KEY.

CLI reference

slidesonnet build                          # build video + SRT subtitles
slidesonnet build --tts piper              # override TTS backend
slidesonnet build --no-srt                 # build without generating subtitles
slidesonnet build --dry-run                # show what would be built (no TTS/FFmpeg)
slidesonnet preview                        # quick build with local Piper TTS
slidesonnet subtitles                      # regenerate SRT from cached audio
slidesonnet preview-slide slides.md 3       # play one slide's audio
slidesonnet preview-slide slides.md 3 -p slidesonnet.yaml  # with playlist config
slidesonnet init md myproject               # MARP Markdown project
slidesonnet init tex myproject              # Beamer LaTeX project
slidesonnet list                           # list slides with cache status per slide
slidesonnet utterances                     # export narration text for proofreading
slidesonnet clean                          # clean cache (keeps API audio by default)
slidesonnet doctor                         # check installed dependencies

Incremental builds

TTS audio is cached by content hash of the narration text, not by slide number. This means:

• No changes → entire build is skipped
• Edit one slide → only that slide's audio is re-synthesized
• Insert a slide → existing slides hit the cache, only the new slide triggers TTS
• Change voice preset → affected slides rebuild (voice is part of the hash)

Use --dry-run (or -n) to see what a build would do without making any API calls:

$ slidesonnet build --dry-run
8 narrated slides: 5 cached, 3 need TTS (~1,200 characters via elevenlabs)

This is especially useful before ElevenLabs builds to estimate API usage and cost.

Build artifacts live in cache/ next to the playlist file. Add it to .gitignore.

Subtitles

Every build automatically generates an SRT subtitle file alongside the video (e.g., my-course.srt next to my-course.mp4). The subtitles use the original narration text (before pronunciation substitutions) and are timed to match the audio.

Long narrations are split into subtitle-sized chunks at sentence boundaries, then clause boundaries, then word boundaries — each chunk timed proportionally by character count.

Use the SRT file as a starting point for translation or editing with any subtitle tool. To skip generation, pass --no-srt. To regenerate from cache without rebuilding:

slidesonnet subtitles

Project layout

my-course/
├── slidesonnet.yaml           # playlist + config
├── pronunciation/
│   └── cs-terms.md
├── 01-intro/slides.md        # MARP module
├── 02-proofs/slides.tex      # Beamer module
├── animations/euler.mp4      # video module
├── .env                      # API keys (gitignored)
├── my-course.mp4             # final output video
├── my-course.srt             # auto-generated subtitles
├── cache/                    # build artifacts (gitignored)
│   ├── audio/                # TTS cache (content-addressed)
│   ├── 01-intro/
│   │   ├── slides/           # extracted PNGs + manifest
│   │   ├── utterances/       # text sent to TTS (for debugging)
│   │   └── segments/         # per-slide video segments
│   └── .doit.db
└── .gitignore

Development

git clone https://github.com/avivz/slideSonnet.git
cd slideSonnet
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[piper,dev]"

make test-unit     # unit tests only (fast, no external tools)
make test          # all tests (requires ffmpeg, marp, latexmk, piper)
make lint          # ruff check + format
make typecheck     # mypy --strict

License

MIT