A cloned voice. A locked dialect. No drift into Fus'ha. No drift into other accents. Listen below.
Listen in order. The human reference. The first attempt that fell apart. The current model — same voice, locked into Jordanian.
Native Jordanian speaker, conversational register. This is the target identity.
Same prompt as the next card. Standard speech architectures aren't built for Arabic dialect — what came out was unstable sound, no real voice identity.
Three takes from the current model. Same voice. Same prompt. Listen to all three.
Every dialect has its own rhythm, emotion, and identity. Most TTS systems collapse all of them into a single Fus'ha-flavored output that sounds nothing like how people actually speak. We're building a voice engine that respects that distinction.
Jordanian is locked. Saudi Najdi, Khaleeji, Egyptian, Levantine, Palestinian, and Emirati are next — same architecture, dialect-specific data.