Alamara

Change the line.
Keep the shot.

Alamara re-syncs an actor's mouth to a new line of dialogue. Send the shot and the line, recorded or typed, in the same language or another, and get the same shot back saying it.

Silent in the original 01 / 09
She says nothing in the original. Here she is given a line, spoken by a computer voice.

Every clip went through Alamara once, with no retouching. The originals play without sound.

Only the mouth is redrawn.

Alamara lays the new mouth back onto your own frames. Eyes, hair, grain and background stay your original picture, at the same size, frame rate and length.

3%of the frame redrawn, on average, in the shot above
6%in the frame where the most changes
A frame from the demo shot in grey, with only the mouth and jaw in colour and outlined in red.
One frame of the shot above. Colour marks what Alamara redrew. Grey is untouched.

Four ways to give it the line

The shot is yours. The line can come from a microphone, a file, a keyboard, or the shot itself.

Record it

Say the line on this page while the shot plays, as many times as it takes.

Upload it

A recording of the line from an ADR session, a voice actor or anywhere else.

Type it

In any of 31 languages. A computer voice says it in the voice of your clip, or of a sample you give it, and checks its own reading before it is used.

Translate it

Type nothing. Alamara hears the line said in your clip, puts it into the language you pick and says that.

Hearing, translating and speaking all run on our own machines. Your dialogue is not sent to anyone else's AI service.

How it works

  1. Send the shot and the line

    One continuous shot with the speaker's face in view, and the new line in any of the four ways above.

  2. The mouth is re-synced

    A video model redraws the mouth and jaw to match the new audio.

  3. The take comes back on your plate

    Only what changed is taken from the model. It is laid onto your original frames, and it keeps the actor's own skin and beard.

What it handles today

Alamara is new. This is what we have tested, including what does not work yet.

Works

  • A face turned toward the camera, straight on or at a slight angle
  • A closed mouth that has to start speaking
  • Beards and stubble
  • Recorded and typed lines in English, Hindi and Spanish
  • A clip's own English line, translated into Hindi and Spanish
  • A voice copied from a clip with music under the dialogue: the music is taken out first
  • One continuous shot of up to 173 seconds. The trial takes 10 seconds; longer shots go through the API
  • A clip with cuts in it: each shot is re-synced on its own
  • A scene that cuts between people: the listener's shots are found and come back exactly as shot
  • Two people in one shot, with a line each
  • A prop that passes in front of the mouth: it stays in front
  • Something that hides the whole face for a moment: those frames come back exactly as shot

Not yet

  • An original where the mouth is wide open: shouting, screaming, laughing
  • A head that turns fast or far to the side
  • A hand or hair crossing the mouth: built the same way as for a prop, not yet tried on real footage
  • Very sharp 4K close-ups: the new mouth can come back softer than the rest
  • Shots longer than three minutes: built for up to ten, tested to 173 seconds

Pricing

The trial is open now. Paid takes open next: leave an address and we will write once, when they do.

Open now

Trial

$0

No account, no card

  • 3 takes a day
  • Clips up to 10 seconds, up to 1080p
  • A watermark on every take
  • Files deleted after 7 days
Try it now
Opening next

Pay per take

$0.10

per second of footage. A 6-second line is $0.60.

  • No subscription
  • No watermark
  • Longer clips
  • Commercial use
Opening next

Studio

Custom

For a film, a series or a campaign

  • Batches of shots in one order
  • A person checks every take before delivery
  • An API key for your pipeline
  • Invoiced, per shot or per project

For pipelines: the API

Whatever the form below does, a program can do: send a shot and a line, ask how it is going, download the take. Takes made with a key carry no watermark and are not held to the trial's length.

Keys are given out by hand while Alamara is new. Each has its own limits: how long a clip, how many takes at once, how much footage a day. Leave an address and say what you would use it for.

# send a shot and a line
curl -H "Authorization: Bearer $ALAMARA_KEY" \
     -F clip=@shot.mp4 -F consent=yes \
     -F text="Buenos días" -F language=es \
     https://mohit1007--alamara-site.us-east.modal.run/api/takes
{"id": "k3Xw9", "state": "queued", "seconds": 6.0}

# the line can also be a recording:  -F line=@line.wav
# or the shot's own line, translated: -F translate=yes -F language=hi
# a clean sample of the voice to copy: -F sample=@voice.wav
# stretches to leave exactly as shot:  -F leave=3.2-7.5,12-14
# or have the listener's shots found:  -F leave=listeners

# ask until it is done, then download
curl -H "Authorization: Bearer $ALAMARA_KEY" \
     https://mohit1007--alamara-site.us-east.modal.run/api/takes/k3Xw9
{"id": "k3Xw9", "state": "done", "took": 214,
 "line": "Buenos días",
 "result": "/api/takes/k3Xw9/result.mp4"}

# or take it back: the GPU stops working on it
curl -X DELETE -H "Authorization: Bearer $ALAMARA_KEY" \
     https://mohit1007--alamara-site.us-east.modal.run/api/takes/k3Xw9

Try it on your clip

Free, with no account. You get a private page that shows your take when it is ready.

  • A person checks each requestNothing is generated until someone has looked at the clip and the line.
  • Trial takes carry a watermark"AI lip sync" is burned into the corner of the picture.
  • Your files are deleted after 7 daysThe clip, the line and the take.
  • No public figures without their agreementRequests that put words in a well-known person's mouth are declined.
01

One shot, 1 to 10 seconds, up to 1080p, with the speaker's face in view.

Drop a video here, or choose a fileMP4, MOV or WebM, up to 100 MB
02
The new line

Say the line as it should be said. It has to fit inside the clip.

Drop a recording here, or choose a fileWAV, MP3 or M4A
or type it

A computer voice says the line in the voice it is given. Best is a sample: 5 to 15 seconds of the person speaking, with no music or other sound. Without one it copies the voice from your clip, with music and noise taken out first, so send the clip with its sound. A recording of the line, if you add one, is used instead of all this.

03
Your say-so

Alamara changes what a real person appears to say.

Optional. We use it only to reach you about this take.