Route 2, the Arabic default

Record the line, measure it, stage for the mouth, generate a sound-off plate at the line's beats, check it free, fit the lips with Sync-3, and ship the exact take.

The route is [best available, not yet validated]; its use as the Arabic default is decided.

1. The recording comes first

A speaking shot waits for its approved recording.

• Record and approve the line (see Voice and Arabic dialects → Voice takes and the delivered file). Lock it: WAV, 48 kHz, 24-bit, mono, with 200 ms of silence at the head and 400 ms at the tail. The locked take never changes. Every copy in this section is made from it and never replaces it.

• Measure it from word timestamps (a transcription tool that gives them, such as ElevenLabs Scribe): the speech onset, the spoken span from first to last word to 0.1 s, every pause inside the line, and the emphatic beat.

• The line must fit the shot at about 2.5–3 words a second, leaving room for reactions. If it doesn't, split the shot, or trim lines once the client agrees; never compress the speech. When an exchange has more beats than a single generation can carry, split it, and assemble the scene one line at a time, each from its strongest take.

2. Stage for lip sync

• One speaker on camera per generation. Everyone else in frame says nothing and has an action.

• Medium close-up or closer, the mouth unobstructed and evenly lit.

• No hand action near the speaking face. Any hand action moves to a separate insert in which nobody speaks.

• The camera moves gently through the line (see Speaking shots and lip sync → The camera during speech).

• Two people talking get a moving master that nobody speaks in, plus one speaking shot each (see Speaking shots and lip sync → Harder speaking shots).

• No start still unless the plan names a start-frame case. A speaking shot that opens on a frame opens on closed lips and a relaxed face (see Video with Seedance → The six start-frame cases).

3. The sound-off speaking plate

• Settings: omni-reference (see Video with Seedance → Platform facts and settings), audio off; references: the character sheet, the identity close-up, the location plate (the approved still of the place; the speaking plate is the video you are making).

• "Sound-off" means no audio, never a closed mouth. Through every stretch of speech the mouth keeps moving, at the take's measured beats, and it rests closed in each pause, before the first word and after the last. Never call the plate "silent": that word brings back a closed mouth. Sync-3 keeps the plate's rhythm and only adds generic lips to a closed mouth, so a closed or murmuring plate fails.

• The written onset comes after the head handle (the 0.5–1 s the clip runs before the action you'll use): about 0.5–1 s in for a line that starts at once; at least 1 s in when the beat needs held silence, and then at least 1 s of closed-mouth silence after the last word.

• Plate length = onset + spoken span + tail (at least 0.4 s, or at least 1 s for silence framing), rounded up to whole seconds. An onset of 0.5 s, a span of 6.0 s and a tail of 0.4 s make 6.9 s, so the plate is 7 s.

• The prompt (Template C; see Speaking shots and lip sync → Templates C, E and F):

  • timed speaking beats from the word timestamps ("1.0–2.6 s she speaks, slow and certain; 2.6–3.2 s she stops, lips closed, eyes on the door");

  • one manner line, and the lips closed in every pause;

  • the line's meaning written as behaviour, never quoted, because readable text is taken as speech;

  • no mouth mechanics: no lip closures or jaw drops written;

  • one speaker, with an eyeline line (see Performance → The eyeline); every listener given a line saying they speak no words, plus a listening task (see Performance → Listening);

  • no voice description beyond the manner line, because Seedance voices nothing on this plate.

• Drafts: 480p in Draft Mode (see Video with Seedance → Draft and finish), then the approved draft finished at 1080p (the 1080p finish). A sample shot gets two drafts. After that, each batch makes one new take of each remaining shot (see Planning, budgets and takes → The generation rhythm). Finish only after a video-analysis report on the draft (see Planning, budgets and takes → Draft small, then analyse). A plain 480p job made outside Draft Mode can't be finished. Set the finish's bitrate to high where the platform offers it: Sync re-encodes the clip, though the gain is not yet known.

4. The draft check and the first-plates check, both free

• The draft check, on the 480p draft against the take's timestamps: the mouth opens within about 5 frames (0.21 s at 24 fps) of each written onset, rests in each written pause, and closes after the last word. A spoken stretch in which fewer than a quarter of the frames show the mouth moving fails at once. A late onset or a murmuring pause is a warning. These are working numbers: log every offset you measure.

• How: with the plate-length file (step 5) laid under the draft, count the frames from each written onset until the lips part. To judge a whole spoken stretch at once, look at the mouth on a contact sheet and count the tiles whose mouth shape changes (steps below).

• The first-plates check, on the project's first speaking plates, in the sample batch: with the sound off, does the mouth move through every spoken stretch while the camera moves?

• If the first-plates check fails on both drafts, write down the diagnosis and decide at your next batch review: rewrite the beats (the usual cause), or move those shots to Route 1's audio-driven form and write that choice in the plan (see Speaking shots and lip sync → Route 1, for close-up hero lines). Nothing switches by itself.

5. Sync-3 on a plate-length file

• Build the plate-length WAV, a copy made from the locked take: room tone across the plate's full length, and the take placed so that its first word lands on the plate's written onset. Place it by the first word, not by the file's start, which carries the 200 ms pad. Export 48 kHz, 24-bit, mono, exactly the plate's length.

• How: build it to the plate's exact length, room tone first, then the take slid until its first word sits on the onset (steps below). If speech would run past the end, the plate is too short.

• The room tone is a quiet, speech-free bed that matches the scene: a designed ambience (see Sound design and mix → Designed effects with ElevenLabs) or a neighbouring shot's own sound. This is a starting practice; no source names one.

• Check the file: its length equals the plate's (steps below), and the first word sits on the onset (word timestamps of the file).

• Run Sync-3 with the 1080p finish as the video and the plate-length file as the audio, and set the sync mode to cut_off yourself: cut_off ends the output at the shorter of the two inputs, while Higgsfield's default, bounce, plays the video forward and back to fill a longer audio. With equal lengths, the output keeps the whole plate, handles included. One speaker per pass.

• The first time you run Sync on a project, confirm:

  • that the job really ran Sync-3;

  • that it took the 48 kHz, 24-bit WAV (if not, give Sync a 320 kbps MP3 copy; the WAV still ships);

  • that the closed mouth holds in the room-tone stretches;

  • the output's resolution and frame rate (Sync may upscale, and it prices on a 25 fps basis);

  • how the price grows with length.

• If the closed mouth breaks in the silences, trim the plate to its in and out points before Sync, and trim the take to match. That's a cut, not a pixel edit.

6. The exact take ships

• The Sync output is the picture. Its own sound is switched off in the timeline, kept and linked: it is a re-encode and would double the dialogue.

• The placed voice, with the dry take kept underneath (decided): the DIALOGUE track carries the line's placed copy (the take set into the room the shot shows), and the untouched dry take sits beneath it on DIALOGUE ALT 1, switched off (see Sound design and mix → The placed voice). Both start where the first word lands on the written onset, at the same offset as the plate-length file. Lip sync always uses the dry take, never the placed copy.

• The plate's own finish sits beneath the Sync output as an earlier take, switched off, with the earlier takes below it (see Editing → The editable timeline).

• Mark the shot's dialogue window in your cut list. Never retime inside lip-synced speech.

• Seedance's ambience and the designed effects go on their own tracks.

Do it yourself

Find the mouth onsets frame by frame (Premiere or DaVinci Resolve), for step 4

1. Lay the plate-length file (step 5) under the draft in Premiere or DaVinci Resolve.

2. Go to each written onset and step with the arrow keys, counting frames until the lips part.

Make a contact sheet of the mouth (ffmpeg), for step 4

1. Set the crop to the mouth box in the draft's own pixels, and export the stretch as one contact sheet:

ffmpeg -ss 0.5 -t 1.8 -i draft.mp4 -vf "crop=160:120:400:260,tile=9x5" -frames:v 1 mouth.png

1. Keep -ss and -t before -i, so the stretch is cut before the tiles are laid out; placed after it, the command writes nothing.

2. The 45 tiles hold just under 1.9 s at 24 fps; for a longer stretch, run it again with a later -ss.

3. Count the tiles whose mouth shape differs from the tile before.

Make the plate-length WAV (Premiere, Resolve or Audition), for step 5

1. In Premiere, Resolve or Audition, make a sequence exactly the plate's length.

2. Lay the room tone across it.

3. Slide the take until its first word sits on the onset; if the head pad is longer than the onset, trim the pad instead.

4. Export 48 kHz, 24-bit, mono, exactly the plate's length.

Read the file's length (ffprobe), for step 5

1. Read the plate-length file's duration, and check that it equals the plate's:

ffprobe -hide_banner -loglevel error -of default=nw=1:nk=1 -show_entries format=duration plate.wav

Sources

• Sync, lip-sync models (a face that already speaks; the plate's own rhythm kept); Ruairi Robinson's Boots (shots) (a performing picture first, lip sync after); Flawless TrueSync on Watch the Skies (New Atlas); Higgsfield's Lipsync Studio post.