ElevenLabs models, same voice, same words
All four rendered with Rami at speed 0.9. Nothing in the app has changed.
What to listen for
The ق. Palestinian drops it to a glottal stop, so قَهْوَة should be ’ahwe, not
qahwa. ElevenLabs does not differentiate Arabic by dialect at the model
level — the accent comes from the voice clone, so switching model should change
the delivery but not the accent. Worth confirming that is what you hear.
| Model | قَهْوَة | الْقَهْوَة عَ الطَّاوْلَة |
|---|---|---|
| multilingual_v2shipping | ||
| v3 | ||
| flash_v2_5 | ||
| turbo_v2_5 |
For reference — what the app plays today
| Nour · Hakim | the current default | |
| Rami · ElevenLabs | trimmed and encoded as the app ships it |
Measured, not assumed.
Both v2 and v3 honour
v3 drops two controls. The API reports
The character limit is irrelevant either way. v3 caps a request at 5,000 characters and v2 at 10,000; the longest text in this app is 30. What matters more is that v3 covers 74 languages against v2's 29 — no use for Arabic alone, but it is where ElevenLabs is putting its effort.
Re-rendering costs almost nothing. 363 clips is about 5,475 characters and you have used 6,391 of 131,000. Switching Rami and Salma to another model is one constant in
speed, but not equally: at 0.7 the sentence runs
2.51s on v2 and 1.84s on v3, so v3 compresses the range and will not go as slow.
v2 rejects speed 1.3 outright; v3 accepts it. Since this is a
learning app and slower is the point, v2 currently gives more room at the slow end.
v3 drops two controls. The API reports
can_use_style: false and
can_use_speaker_boost: false for v3, against true for both
on multilingual_v2. Not used here yet, but it is a one-way door if you ever want them.
The character limit is irrelevant either way. v3 caps a request at 5,000 characters and v2 at 10,000; the longest text in this app is 30. What matters more is that v3 covers 74 languages against v2's 29 — no use for Arabic alone, but it is where ElevenLabs is putting its effort.
Re-rendering costs almost nothing. 363 clips is about 5,475 characters and you have used 6,391 of 131,000. Switching Rami and Salma to another model is one constant in
tools/eleven_render.py plus a delete of their clip folders.