Voice and Video

Voice is opt-in. A deck with a voice/ directory can synthesize its script with a local text-to-speech engine, rehearse to that audio, and bake an MP4. A deck without voice/ is unaffected, and what fails its lint does not change.

Voice and video are derived from the script. The script knows nothing about either.

script.md → Deck → Cue → Synth → Timeline → schedule
                                      ├→ rehearse  (the browser drives itself)
                                      └→ video     (frames → ffmpeg)

Only paragraphs are spoken

dekc cues

This prints the spoken cues for every beat and needs no engine. Blockquotes, lists, code, and tables are dropped; inline emphasis, links, and code spans are flattened to text. A beat with visible content but no paragraph is reported as DEK042 alongside the cues, so you notice it while you are still writing.

Setup

Add a [voice] table to dek.toml, and dekc new copies it into each new deck as voice/voice.toml. For an existing deck, write the file yourself.

# decks/<deck>/voice/voice.toml
engine  = "voicevox"
speaker = "ずんだもん/ノーマル"
speed   = 1
pause   = { sentence = 350, beat = 700 }

dekc voice needs a running VOICEVOX-compatible engine; rehearse and video need only the Timeline it writes. If no engine is found, only the command that needs one fails, and the hint tells you how to install one.

• VOICEVOX at port 50021: voicevox.hiroshiba.jp, or docker run --rm -p 127.0.0.1:50021:50021 voicevox/voicevox_engine:cpu-latest
• AivisSpeech at port 10101 with engine = "aivis": aivis-project.com, or docker run --rm -p 127.0.0.1:10101:10101 ghcr.io/aivis-project/aivisspeech-engine:cpu-latest
• COEIROINK and SHAREVOX are recognized by name.
• Any compatible engine elsewhere: set engine to its URL, or export DEK_VOICE_URL.

Synthesis

dekc voice
dekc voice speakers
dekc voice say "Hello"
dekc voice dict add dek デック
dekc voice pin

dekc voice synthesizes only the sentences that changed and writes the audio, the per-sentence cache, and timeline.json under .cache/voice/. The dev server does the same on save. An ASCII word missing from voice/dict.toml is DEK040, a warning; add readings with dict add. dekc voice pin copies the master audio and Timeline into voice/pin/, a portable snapshot that survives a cleared cache. While the pin exists, dekc voice and the dev server restore it instead of synthesizing, so edits to the script or to voice.toml do not reach the audio; delete voice/pin/ to synthesize again.

Kana and durations are machine-readable. dek does not default to a cloud TTS.

Rehearsal

dekc rehearse

The dev server starts with the Timeline as its clock and the browser advances beats on time with the audio. Nothing is recorded. Space plays and pauses; the arrow keys and dekc goto seek, and the audio follows. A hand-written timeline.json is enough to rehearse without an engine.

Video

dekc video
dekc video architecture
dekc video --fps 30 --root-dist

A whole deck becomes dist/<deck>.mp4, or <root>/dist/<deck>.mp4 with --root-dist. A single slide goes to .cache/video/<slug>.mp4. Requires Playwright, ffmpeg, and a Timeline; without a Timeline the hint says to run dekc voice first. Alongside the MP4, dek writes .vtt captions, a .chapters.txt chapter list, and a .credits.txt file naming the speaker. Credits are not burned into the picture.

Baking does not replay the talk in real time. Each beat is one held frame plus whatever the transition needs.

When a Timeline exists, dekc ls shows the narrated length next to the word-count estimate, and a large gap from the duration budget is DEK041, a warning. The script has no notation for silent time: a beat with no paragraph passes through the transition and pause.beat, and nothing else. Timing lives in voice.toml, below.

Timing

Rehearsal and video share one schedule. Each screen change leads its first word by lead milliseconds (300 by default), so a transition has settled by the time the voice arrives. A lead longer than the beat before it never fires ahead of that beat's own change. When one beat still feels early or late, adjust that beat in voice/voice.toml; the script stays as it is.

lead = 300                  # the whole deck

[beats.order]               # the whole slide
lead = 600                  # before the slide arrives
pause = 1200                # after its last beat

[beats."order/what"]        # one beat, by id or by 1-based position ("order/2"); "order/0" is the arrival
pause = 1500                # silence after it, instead of pause.beat

Keys name a slide as slug and a beat as slug/beat-id or slug/2, the position the URL hash takes. A slide key frames the slide: its lead runs into the slide's arrival, where the paragraphs before its first ### are spoken, and its pause follows the last beat. A beat key wins over its slide key. dekc mv rewrites the keys along with the slide. A key that matches nothing is DEK043. On save, the dev server re-times the deck from cached clips, and dekc voice pin freezes the timing together with the audio.

The daily four

$EDITOR script.md
dekc
dekc rehearse
dekc video

For a live-only talk, the three commands in Getting Started are all there is.

Next

Let an agent write, check, and fix slides: Working with AI Agents.