DeckTalk

Every picture lands on its word, and you edit the video like a doc.
Before this, one wrong word meant editing, rendering and recording the whole thing again.

DeckTalk built this film from text files, and every reveal in it starts on its word. Watch it with the sound on, or read the whole pitch at decktalk.ai.

Quickstart · Docs · Requirements and costs · Reference card for agents · Changelog
Agents can read llms.txt.

What it is

DeckTalk makes a narrated video from a markdown script and plain HTML slides. Your cloned voice reads the script and comes back with a time for every word, so each reveal starts on the word that introduces it. The script and the slides are text you keep in a repository, so you change a sentence and build again, the way you would change a doc or a codebase.

Four things are different from a video editor, a slide tool, or an animation library.

• Every reveal is cut to the spoken word. A cue names a phrase, not a second, so a reveal follows the narration wherever the voice puts it. decktalk verify measures every reveal in the finished video, and in the sample below each one lands within one frame of its word.
• A changed sentence voices only its own section again. DeckTalk caches every section's narration by its text and voice settings, so an edit spends credits only on the sections that changed. build --only N records only the section you name.
• The whole video is text. You write four files, and an agent can write them too. There is no timeline to drag and no project file a person cannot read.
• Your coding agent already knows how to drive it. decktalk init installs six skills into the project, one each for the script, the slides, the cues, the build, the fixes and a revision, so an agent writes the files, reads the JSON every command prints, and fixes what it finds. The skills lists them.

Who it is for

• Technical tutorials. A viewer watches a walkthrough whose commands and screenshots match the version she just installed. When the library changes, you change the sentence and build again.
• Product demos. A viewer sees the feature narrated over the screen it runs on, with every callout landing on the word that names it.
• Lessons. A viewer follows an explanation in which each idea appears as it is spoken, so nothing is read ahead.

Internal presentations and estimation walkthroughs run the same pipeline. Neither has a shipped example yet.

Make your first video

The first line installs uv, a Python package manager, and then DeckTalk. It brings its own Python, so there is nothing to install before it. Read the script before you run it; that is why it is served from a URL. If you already have uv, uv tool install decktalk does the same thing, and pipx install decktalk works too.

curl -LsSf https://decktalk.ai/install.sh | sh
decktalk install
decktalk init my-lesson && cd my-lesson
decktalk build --no-voice

decktalk install downloads Chromium and ffmpeg one time per machine, and on Linux it asks for sudo. decktalk init writes the starter, a working three-section project with one equation, and nothing in it has to be deleted first. The build without voice needs no account, ends with built, and prints the path of build/out/my-lesson.mp4. Recording runs in real time, so the build takes about a minute and makes a 44 second video.

To hear it in your own voice, copy .env.example to .env, set your ElevenLabs API key and voice id, and run decktalk build. DeckTalk never prints the key. A voiced build of the starter sends about 420 characters, and decktalk narrate --dry-run prices that run before it starts. Every build writes the mp4, SRT and VTT captions, a chapter per section and a transcript page. What spends credits lists the cost of every command.

The quickstart shows the output of each step and what the starter shows. If you try DeckTalk on one section of something you teach, tell me in Issues what stopped you.

What you write

You write four files. One cue id per reveal, such as 1.1script, ties the script, the cues, and the page together.

The narrate stage writes a words file with the start and end of every spoken word. If a cue phrase is not in the spoken words, the build stops and names the cue. Your first deck writes one section in all four files.

The script comes first, the voice gives every word a time, and the cues are where a named phrase meets its picture. Change one sentence, and only that section is voiced and recorded again.

Why the cuts are exact

A browser does not start recording at a known time, so DeckTalk does not use a timer. The recorder covers the page in magenta until the narration starts. The first frame without magenta is narration t=0, on Linux, macOS, and Windows.

decktalk verify then measures every reveal in the finished video. This sample comes from a build without voice of the starter, and one frame lasts 40 ms.

check                 cue       at   chg %   ctl %   offset     a/v  result
1:1.1title           0.70     0.70    1.87    0.00    +20ms   +16ms  changed
2:2.1code            8.93    23.41   10.79    0.00    -10ms   -34ms  changed
3:3.1make            8.21    38.13   10.85    0.00    -10ms    +8ms  changed

The offset column is the time from the cue time to the onset of the reveal, in milliseconds. Verify defines every column and limit.

Requirements and costs

• Software. The one-line installer brings its own Python, on Linux and macOS. On Windows, install with uv or pipx, which needs Python 3.12 or later. decktalk install downloads the rest.
• Accounts. A build without voice needs no account. A voiced build needs an ElevenLabs API key and a voice id.
• Cost. Every ElevenLabs plan can call the API. The free plan has limits for a video you publish.

Requirements and costs lists every download and every command that spends credits.

How it compares

The FAQ has the full comparison. Every build also gives you these parts:

• A free build without voice. Every stage runs with no key, and a click marks each word.
• Cached narration. DeckTalk voices a section again only when it changes, and keeps a recording whose page, words and cues have not moved.
• Captions, chapters and a transcript. Every build writes SRT, VTT, chapters and a transcript page.
• A soundscape. Add music, an ambience bed, and sound effects.
• Loudness. DeckTalk normalizes the mix to -16 LUFS.
• Checked output. record and verify catch bad recordings and late reveals.

Status and support

DeckTalk is alpha, and a minor release can still break things. One person maintains it.

CI runs the unit tests and a full offline build on Linux for every push to main and every pull request. CONTRIBUTING lists the macOS and Windows runs.

• Report bugs, ask questions and show what you made in Issues.
• Report a vulnerability privately. SECURITY.md explains how.
• The changelog lists every release.

Documentation

The docs are at docs.decktalk.ai.

• Install and build a first video: Quickstart
• Write your own section: Your first deck
• Give an agent the whole contract: Reference card for agents
• Understand how it is built: ARCHITECTURE.md and the decision notes

Contributing

To work on DeckTalk itself, start with CONTRIBUTING.md.

Apache-2.0. Made by Jacob Beaudin.