The Master Guide / 10 Voice and lip sync · Voice and Arabic dialects
Cast the dialect, don't prompt it
A voice brings its own accent; a tag, a language code or a style line cannot add one [official guidance]. Choose a voice that already speaks the dialect. On v4 in Arabic this casting rule is [best available, not yet validated].
A voice brings its own accent; a tag, a language code or a style line cannot add one [official guidance]. Choose a voice that already speaks the dialect. On v4 in Arabic this casting rule is [best available, not yet validated].
A library voice's accent tag is set by whoever uploaded it, so the tag proves nothing: every candidate goes to a native ear.
Dialect | Start with | Human fallback |
Modern Standard Arabic, MSA (formal voiceover, documentary, pan-Arab corporate) | An MSA library voice | A professional MSA voiceover artist |
"White" (neutral) Arabic | No engine offers one. Make it by casting plus copy: a mild Gulf or Levantine voice reading pan-Arab wording without strong local markers; a Gulf voice for Gulf-facing work, an Egyptian one for Egypt-facing work | A pan-Arab voiceover artist |
Egyptian, Cairene | An Egyptian library voice | An Egyptian artist for acting-critical lines |
Egyptian, Sa'idi | No AI Sa'idi voice was found in the library (searched 20 Oct 2026). Try the three most suitable Egyptian voices on the Sa'idi copy | Most likely a native Sa'idi actor |
Saudi, Najdi or Hijazi | A Saudi voice matched to the sub-dialect | A native Najdi or Hijazi artist |
Gulf (UAE, Kuwait, Qatar, Bahrain, Oman) | A voice from the target country | A native Gulf artist |
Levantine, Maghrebi, Iraqi, Sudanese, Yemeni | A matching voice, found with the accent filter | A native artist |
Children | No source shows an Arabic child voice working on v4 | An adult native actor performing the child, or a child's read |
Cast in ElevenLabs only; no other speech engine is brought in (decided), because the stack stays lean.
Clones
A Professional Voice Clone (30 minutes or more of a native actor's clean, dry speech) is for a real person's voice, a recurring brand voice, or a role no library voice fits. ElevenLabs runs the verification.
Instant clones and Voice Design are for drafts and fictional casting only. ElevenLabs calls Voice Design less performative on v4, and an instant-cloned Saudi voice scored lowest in the only Arabic benchmark. Never use a one-shot clone for an Arabic final.
Record a prepared read for any clone, covering the dialect's key sounds, not free speech. If a clone made before 28 Sep 2026 sounds wrong on v4, retrain it.
Sources
ElevenLabs, Egyptian accent text to speech and voices; the Habibi benchmark (arXiv).