ElevenLabs settings, tags and pace

Set every value on every call. Pace comes from the words, never from the edit.

In the web app, set the same values by hand for every line (steps below).

Delivery tags

Write them this way [official guidance]:

• in English, in square brackets, used sparingly, placed next to the words whose delivery they shape;

• several descriptors may share one bracket: [calm, slow, low voice];

• describe a quality of the voice; a bare one-word tag can be read as a sound effect;

• tags live only in the engine text. If the transcript shows a tag word spoken aloud, move or reword the tag.

Pace

v4 has no speed control, so pace comes from the wording, the punctuation, ellipses and the tag.

• Never time-stretch, speed up or pitch-shift speech in the edit [official guidance]. The one exception: in a 25 fps TV version, lip-synced dialogue follows its picture by the broadcast speed-up, a plain resample (see Graphics, captions, versions and delivery → 25 fps TV versions by speed-up).

• A line too long for its beat is rewritten shorter, with the same meaning. You approve the new wording, and the client does too if it changes approved copy.

• Measure each voice's Arabic speaking rate from its first approved take; English figures (2.4–2.6 words a second) don't transfer. Plan on-screen lines at about 2.5–3 words a second.

Exchanges and runs of lines

• Several speakers: record the exchange together in v4's dialogue mode (10,000 characters at most), then split it into one file per speaker.

• Consecutive lines of one speaker in one shot are recorded and locked as one take, with the timed pause written in (an ellipsis or a line break); your line ledger lists both lines on one row. Two separate takes would lose the breath between them.

Two characters in dialogue mode

Write the exchange as one scene, because comic timing lives in the gaps [best available, not yet validated].

1. One speaker per turn, in script order, each turn one short thought, with its own delivery tag where the attitude changes ([cutting in, impatient], [deadpan]).

2. An interruption: end the cut-off turn mid-word or mid-phrase with a dash, and start the next turn with a tag that says it cuts in. Two people speaking at once are never generated as one: record each side whole, then slide the second speaker's file earlier in the edit.

3. Room for the laugh: an ellipsis, or a short separate turn with [beat] before the reply, where the joke needs air. Measure the gap on the take and keep it in the cut.

4. Generate the whole exchange twice, and keep both.

5. Split by speaker without losing the timing: from the exchange's word timestamps (Scribe with speaker labels on), cut each speaker's turns into their own file, padded with silence to the exchange's full length, so every file starts at the same zero. Lay them all at one shared start in your cut list; the gaps stay exactly as recorded.

6. A weak turn: regenerate the exchange from that turn on, never the turn alone, so the rhythm holds. You and the native listener judge the exchange as a whole.

An interrupted turn, ending on its dash:

«استنى بس، أنا كنت لسه هقول—»

An AI interview voice that should sound recorded

For a documentary-style film where a generated voice plays an interviewee [best available, not yet validated]:

• Write speech, not copy: short clauses, a restart or two, a small hesitation word the dialect uses, a breath at a change of thought, a sentence that trails off. No ad phrasing.

• Tags for the situation: [thoughtful, unhurried, talking to someone off camera], [remembering, a little emotional]. Stability on the low side, about 0.4.

• The microphone's distance comes from the mix, not the engine (see Sound design and mix → The placed voice): a small room, a chest-microphone sound, over the location's room tone.

• Breaths and small sounds stay in; never strip them as noise.

• The native listener judges whether it sounds like a person talking, not reading. A real interviewee recorded on location always outranks it.

A restart, as an interviewee might say it:

«يعني… هو أنا لما… لما فتحت المحل»

Do it yourself

Make a line in text-to-speech (ElevenLabs web app)

This Guide doesn't name the app's buttons, and their places may move; the values are what matter. On the API, send the same values on every call.

1. Sign in to your own ElevenLabs account, on a paid plan, and open text-to-speech.

2. Choose the anchored voice from your account's own voices, by the name and ID you recorded (see Voice and Arabic dialects → Casting a voice). Never pick it from memory.

3. Choose the model Eleven v4, never v4 Turbo.

4. Set the values in the table above:

   • the language: Arabic (ar) or English (en), if the app offers it;

   • Stability at the line's value: 0.5 to start, or the value its line script gives;

   • Similarity left at its default;

   • the project's alias dictionary, loaded from your pronunciation sheet;

   • the output at 48 kHz WAV or PCM, or the highest PCM rate your plan allows.

5. v4 has no Speed control. If the app shows one, note that at first use and leave pace to the words.

6. Paste the engine text: the tag first, then the line (see Voice and Arabic dialects → Ten complete line scripts). Keep it to one line, or one short group of sentences.

7. Generate it twice, and keep both takes.

8. Write the model, voice, settings and date in your take log, and the engine text in your line ledger.

Sources

• ElevenLabs, What is Eleven v4?, the v4 prompting guide and the text-to-speech API reference.