Eleven v4 Launch: What Changes for Voice Content
ElevenLabs released Eleven v4 and v4 Turbo on September 28, 2026. Here is what the launch notes say and what marketers and founders should test first.
ElevenLabs released two new text-to-speech models on September 28, 2026: Eleven v4 and a faster sibling, Eleven v4 Turbo. According to the company's launch post, v4 is aimed at more expressive speech, and Turbo at low-latency use such as live agents.
If you produce voiceovers, ads or product videos, here is what is new and what we'd test first.
What did ElevenLabs announce?
Based on the company's launch post and changelog entry for September 28, 2026:
- Two models. Eleven v4 for expressive narration, and Eleven v4 Turbo for speed.
- Latency. Coverage of the launch cites a median inference latency of about 100 milliseconds for Turbo and a median time to first speech of about 150 milliseconds. Both are vendor figures, measured on its own setup.
- Languages. Both cover about 90 languages, up from 70 in the previous version.
- Expressive control. The models read tone, pacing and context, and accept inline audio tags in the script, such as [laughing] or [whispering]. SSML break tags are switched off in v4 in favor of these natural-language cues.
- Voice cloning. A voice can be cloned from about 10 seconds of audio. Older Instant and Professional clones need retraining for v4, according to launch coverage.
- Where it runs. Available in ElevenAgents, ElevenCreative and through the API.
The launch post also makes ranking claims. Those are the company's statements, so treat them as marketing until you hear the model on your own script.
Why does it matter for marketers?
Two things stand out.
First, audio tags. Direction like "laughing" or a mood cue inside the script means fewer re-takes and less editing. For short ads, that can be the difference between a voice that sounds read and one that sounds said.
Second, the short clone sample. Ten seconds lowers the barrier to creating a branded voice. It also raises the stakes on consent. Clone only voices you have written permission to use, and keep that permission on file.
What should I test first?
We'd run a 30-minute test before changing any workflow.
- Take one real 30-second script, including your product name and any numbers.
- Generate it with v4 and with your current voice.
- Add two or three audio tags where a human would pause or smile.
- Render it in your two main languages and have a native speaker listen.
- Compare cost and turnaround, not just sound.
If the pronunciation of your brand name is off in a language, fix it with a pronunciation dictionary or spelling before you judge the model.
What to watch next
Pricing, rate limits and the real-world quality in less common languages are the open questions. Our guide on choosing an AI video, voice or image tool covers the rights and cost checks to run before you commit, and the AI Images, Video & Media hub follows new releases in this space.
FAQ
What is Eleven v4?
It is ElevenLabs' text-to-speech model released on September 28, 2026, built for more expressive speech with inline audio tags.
What is the difference between v4 and v4 Turbo?
Turbo is the low-latency version for real-time use, with a median inference latency the company puts at about 100 ms. v4 is the expressive model for narration and content.
Can I clone a voice with 10 seconds of audio?
ElevenLabs says a clone now works from about 10 seconds. You still need permission from the person whose voice it is.
