The Master Guide / 10 Voice and lip sync · Voice and Arabic dialects
Voice takes and the delivered file
Two takes per line, measured, chosen by a native ear, locked, and never touched again. Voiceover is made before the cut and before the music.
Two takes per line, measured, chosen by a native ear, locked, and never touched again. Voiceover is made before the cut and before the music.
Estimate first. List the lines, voices and characters with marks (full marks roughly double a line's characters), the number of generations, and the price in US dollars: $0.08 per 1,000 characters at the API rate (read 19 Oct 2026). If you generate in the web app, also note what the lines spend in ElevenLabs credits, never mixed with Higgsfield credits. Get the estimate approved before generating. The video batch rhythm doesn't change voice: every line keeps at least two generations.
Generate two takes per line with the settings in Voice and Arabic dialects → ElevenLabs settings, tags and pace, and download both (steps below). Name voiceover and on-screen dialogue apart, so the two never mix on the timeline.
An optional cheap first pass: check only the transcript's numbers and proper nouns. If one is wrong, fix the text and regenerate at once when the approved estimate covers the retry; otherwise get the extra cost approved first.
Measure every take (see Voice and Arabic dialects → Measured checks and the back-check) and place it on the assembly or in its shot window, the span of frames your shot plan gives that shot (see Planning, budgets and takes → Timing, grouping and trimming).
The native ear decides (see Voice and Arabic dialects → The native reviewer). A take may be accepted as "some flaws, but good"; record the verdict in the listener's own words.
Lock the chosen take and never change it. Every derived copy is made from it and never replaces it: the plate-length file for lip sync (a copy laid on room tone at the speaking shot's exact length, its first word on the shot's written onset; see Speaking shots and lip sync → Route 2, the Arabic default), or a reference padded with room tone to at least 2 s.
The delivered file
Your internal master:
Item | Spec |
Source | Lossless from the engine (48 kHz WAV or PCM); note the source rate in your take log |
Master | WAV, 48 kHz, 24-bit, mono, dry: no music, no ambience. Resampled once, with a high-quality resampler |
Per file | One line, one speaker |
Trim and pad | Trim the engine's tail click; 200 ms of digital silence at the head and 400 ms at the tail; record both the raw and the padded lengths |
Level | One gain per voice per scene, recorded: about −20 LUFS integrated, with a −3 dBTP true-peak ceiling. LUFS (loudness units relative to full scale) is the unit loudness is measured in; dBTP (decibels true peak) is the unit of the true peak, the highest level the waveform reaches, including between samples. For lines under about 3 s, use true peak plus the scene gain. Intended whispers and shouts are exempt. Delivery loudness is set in finishing (see Sound design and mix → Loudness: −23 TV, −16 online, −14 short social) |
MP3 | A review copy only |
To set the level, read the take's integrated loudness and true peak on an EBU R128 loudness meter (a meter built to the European broadcast loudness standard; steps below).
The clean take ships
The locked voiceover stays close and dry. An on-screen line plays as its placed copy, with the dry take muted beneath it on DIALOGUE ALT 1 (see Sound design and mix → The placed voice); every other take sits muted beneath both (see Editing → The editable timeline). Audio a video model returns is never the shipped dialogue. The one exception is a re-voiced read outside Arabic, logged as such (see Speaking shots and lip sync → The re-voiced read); no Arabic re-voiced read ships in any version, drafts, animatics and social cut-downs included (decided).
Voiceover to picture
Make it as soon as the voice is anchored, before the cut and before music, and cut the assembly to the real voiceover [official guidance].
Split the script into the lines your shot list places on picture beats.
Each line starts on its beat and ends before the next key visual or the packshot.
Before the last line and the logo, leave a gap of about 0.3–0.5 s, unless the treatment asks for a tight button.
Keep the natural breaths.
Captions: see Graphics, captions, versions and delivery → Captions, Arabic included.
Do it yourself
Download the two takes (ElevenLabs web app)
Download each take in the lossless format you set: 48 kHz WAV, or the highest PCM rate your plan allows. Never use an MP3 as the master: an MP3 saved as WAV has lost quality for good. An MP3 is a review copy only.
Keep both takes of every line.
Name voiceover and on-screen dialogue apart, so the two never mix on the timeline.
Note each file's source rate in your take log, then trim, pad and level the master as the table above says.
Measure a take's loudness and true peak (Premiere, Resolve or ffmpeg)
In Premiere, put the Loudness Meter effect on the take; in Resolve, read the Fairlight meters. Where to find Premiere's meter: Sound design and mix → Loudness: −23 TV, −16 online, −14 short social → Do it yourself.
Or read both with ffmpeg:
ffmpeg -i take.wav -af ebur128=peak=true -f null -
Sources
ElevenLabs, v4 prompting guide (voiceover before the cut).