Diffcut
Get early access
EditingTrims, keyframes, transitions, export.CaptionsTranscribed on your Mac, word by word.Voiceover28 voices, with pauses where you write them.AI musicMusic and effects from a prompt.CaptureA built-in browser for shots and recordings.Ask ClaudeYour own Claude proposes edits you review.ShortsVertical clips from a long video.Demo videosA first cut from a browser recording.All features
In the app

A voice for every draft.

Write the script, pick a voice, press Generate. Diffcut makes the take on your Mac, levels it, marks it, and puts it on the voice track.

Get early accessMeet the voices

Directions in [brackets].

Anything in square brackets is a direction. A pause becomes real silence in the take; anything else is a note that’s never read aloud.

The coast was quiet the year I came back.
[pause]
Most mornings, the fog sat on the water.
[warmly] And then it lifted.
  • [beat]0.6 s
  • [pause]1 s
  • [long pause]2 s
  • [very long pause]3 s
  • [pause 1.5s]As written, up to 10 s

28 voices, on your Mac.

American and British, female and male. Each runs on the Mac’s GPU, so nothing you write leaves it.

  • American, femaleAlloy, Aoede, Bella, Heart, Jessica, Kore, Nicole, Nova, River, Sarah, Sky
  • American, maleAdam, Echo, Eric, Fenrir, Liam, Michael, Onyx, Puck, Santa
  • British, femaleAlice, Emma, Isabella, Lily
  • British, maleDaniel, Fable, George, Lewis
~2 s
from Generate to a playable line in Diffcut, on an M5
−16 LUFS
every take levelled, at most −1 dBTP, 48 kHz
Every take
watermarked, then checked before you hear it

A watermark on every take.

Each take carries an inaudible AudioSeal watermark with a 32-bit payload. Diffcut reads it back before the take reaches your timeline, and there is no switch to turn it off.

Questions

Which languages can the voices speak?

English, in American and British accents. More languages are planned with a second voice model.

What happens to a direction like [softly]?

It stays in the script as a note for you and is never read aloud. Today’s voice model can’t act a direction out, so only pauses change the take.

Can I turn the watermark off?

No. Every take is marked, and the mark is read back before you hear it. A take whose mark doesn’t verify is made again, at most three times, and never shown unmarked.

Is my script sent anywhere?

No. The voice runs on your Mac’s GPU, so the script and the audio stay on the machine.

More of Diffcut

Make the cut.

Diffcut is free, open source and early: the first Mac build isn’t out yet. Leave your email and we’ll write once, when it’s ready.