Casting a voice

Shortlist three voices per role, expose the dialect, measure, and let a native listener of that dialect choose. Then anchor the voice and never change it.

Casting steps

Before the anchors presentation, and internal:

1. Shortlist three voices per role from the Voice Library: filter by Arabic, the accent, gender and age, and match the treatment's casting notes. On 15 Oct 2026 the Arabic plus Egyptian filter returned 57 voices (14 female, 43 male). Filter Saudi voices the same way, by the sub-dialect; no count for them is recorded.

2. Two dialect-exposing lines per voice, two generations each. The lines carry the dialect's marker sounds and words: for Cairene, jīm said as a hard g and qāf as a glottal stop; for Najdi, qāf said as g and the feminine ending as the sub-dialect says it; for Hijazi and Gulf, their own everyday words (below). The native reviewer confirms the marker words before you make the samples.

3. Measure every sample (see Voice and Arabic dialects → Measured checks and the back-check): the words against the approved line, the length, words per second, pauses of 0.8 s or more, integrated loudness and peak. Measurements narrow the field; they never judge an accent.

4. A native listener of that dialect picks (see Voice and Arabic dialects → The native reviewer).

5. Anchor the voice. Add the chosen library voice to your account's own voices, because a shared voice can be withdrawn by its owner. Record the voice's name, ID and approved sample, and the model, Stability, Similarity, dictionary and the shared voice's notice period. Re-read them before every generation; never pick a voice from memory.

6. Never change model, voice or settings mid-project [official guidance]. Consistency breaks first at a switch of source or a small change of speed or volume.

On the client's page: voices are cast before the anchors presentation, which shows only the chosen voice (see Client presentations → The anchors presentation, in its twelve sections); a speaking shot waits for its approved recording.

The blind casting listen

An optional casting step: a way for the listener to choose among the shortlist on the film's real lines. Use it where you or the native reviewer want the shortlist heard blind; otherwise the pick by ear stands.

1. Lines: four real lines per dialect, about 60 characters each:

   • a plain statement;

   • a number or price;

   • a brand or English term;

   • an emotional or comic line.

2. Voices: the three shortlisted voices per role.

3. Takes: two per voice per line, on full vowel marks. Level every take to equal loudness and give each a random code. Add one hidden duplicate per dialect: the same file under two codes.

4. Listening: each line's candidates played together, in a random order that changes per line. Rate each take 1–5 on dialect authenticity, pronunciation and naturalness, plus yes or no on "as good as our English ElevenLabs reads". A native human read that already exists can be mixed in as an anchor; never pay for one. Add a second native listener where one is free without delay.

5. Measured beside each take: the back-transcription (plus Gemini 3.5 Transcribe for Egyptian), wrong content words, length, words per second, pauses of 0.8 s or more, peaks.

6. The scoring rule, written before anyone listens:

   • a voice's score on a line is the mean of the three ratings of its better take, and its overall score is the mean over the four lines;

   • a 2-point gap between a voice's two takes is logged as inconsistency;

   • a 2-point gap between the hidden duplicate's two codes means that listener needs a second listener before deciding;

   • only a voice with no wrong content word and a pass on the English-quality bar is eligible;

   • within 0.3 points, fewer inconsistencies wins, then the voice with more pronunciation controls.

7. Cost: about $0.30 (US dollars) per dialect on your ElevenLabs account (read 19 Oct 2026), and 45–60 minutes of listening per dialect.

A spreadsheet holds the codes, the ratings and the rule. A voice that no listener accepts sends the role to the fallback (see Voice and Arabic dialects → The voice route).

Checks at first use

Four free checks on your first voice job:

• which ElevenLabs plan allows 48 kHz WAV;

• whether v4 honours the Arabic language code;

• whether aliases and IPA (the International Phonetic Alphabet) work inside Arabic text;

• the shared voice's notice period.

Sources

• ElevenLabs, v4 prompting guide (consistency across sources and settings).