The argument in one line.
Eleven v4 reads a script for context the way a voice actor does, while Eleven v4 Turbo trades a little of that nuance for about 100ms latency so the same expressive model can run inside live voice agents.
Read if. Skip if.
- Anyone building with ElevenLabs who needs to choose between the standard v4 model and the low latency Turbo version for a specific use case.
- Developers building voice agents, audiobooks, or long narrated content who have run into voices drifting or losing consistency across a project.
- Creators working in Japanese, Brazilian Portuguese, Mandarin, or Cantonese who need better pronunciation and delivery than earlier ElevenLabs models offered.
- You don't use ElevenLabs or a comparable text to speech API. This is a feature rundown of one company's new model, not a general voice AI explainer.
- You're looking for pricing, benchmarks, or a side by side comparison against competitors. None of that is covered here.
The full version, fast.
ElevenLabs releases Eleven v4, a text to speech model built on a new architecture that reads a script's context the way a voice actor would, tracking who's speaking, what just happened, and how a line should land, so multi-speaker dialogue gets real overlap and timing even without audio tags. Eleven v4 Turbo is the same research tuned for about 100ms latency so live voice agents can use it in real time. Voice cloning holds up better across long-form regenerations, context stitching keeps a regenerated line consistent with the chapter around it, and language coverage improves most in Japanese, Brazilian Portuguese, Mandarin, and Cantonese. It's live now, free to try for 11 days.
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
01 · Cold open: what's new in v4
ElevenLabs states the pitch directly: v4 is the most expressive TTS model yet, built on a new architecture that reads tone, pacing, emotion, character, and context, with a low-latency Turbo sibling alongside it.
02 · Two models: v4 vs v4 Turbo
Both models share the same research; the only difference is latency. v4 targets produced content (narration, character work, dubbing) across ElevenCreative, ElevenAgents, and the API. Turbo targets real-time voice agents at about 100ms median latency.
03 · Script-aware delivery and audio tags
v4 reads a script the way a voice actor would, tracking who's speaking and what just happened. A scripted two-person exchange demonstrates overlap and timing without relying on audio tags, then the video shows how stacked audio tags direct delivery on top of that.
04 · Voice cloning stability and context stitching
Instant and Professional Voice Clones work identically across v4 and Turbo, but v4 holds speaker identity far more reliably across regenerations and long-form work. Existing PVCs need to be manually retrained on v4. Context stitching then keeps a regenerated line consistent with the rest of its audiobook chapter.
05 · Multilingual gains, Turbo for agents, and how to try it
v4 covers 90+ languages with the biggest gains in Japanese, Brazilian Portuguese, Mandarin, and Cantonese, plus better IPA support. The video closes by pointing agent builders back to Turbo for its bidirectional streaming and real-time latency, then directs viewers to sign up and try v4 free for 11 days.
Lines worth screenshotting.
- Eleven v4 and Eleven v4 Turbo share the same underlying research and differ only in latency, not in expressive range.
- Turbo runs at roughly 100ms median latency, fast enough for live voice agents while keeping v4's full expressiveness.
- Eleven v4 infers tone and pacing from punctuation and sentence length alone, so audio tags only need to override what the text already implies.
- Multi-speaker dialogue in v4 is generated with awareness of the full exchange, not as isolated lines stitched together afterward, so overlap and timing feel like a real conversation.
- Voice identity in v4 holds stable across regenerations and long-form work, a known weak point in earlier ElevenLabs models.
- A Professional Voice Clone trained on an older model has to be manually retrained on v4 before it gets the new model's stability gains.
- Context stitching means regenerating one line inside an audiobook chapter now accounts for the lines before and after it, not just that line in isolation.
- The on-camera host in this video is itself an Eleven v4 voice clone driving an AI avatar built inside ElevenCreative, making the demo its own proof.
What actually changed under the hood in v4
Eleven v4's gains come from reading a script for context before generating it, not from adding more manual control, and that shift is worth understanding even outside ElevenLabs.
- Context-aware generation beats manual tagging: a model that infers tone from punctuation and prior lines needs less hand-holding than one that requires explicit direction for everything.
- Multi-speaker dialogue sounds real when it's generated as a full exchange rather than individual lines assembled after the fact, since timing and overlap depend on both sides.
- Voice consistency over long content is a harder problem than a single good sample. It shows up specifically in regenerations, dialogue turns, and full-length projects, not in short clips.
- Splitting a model into a produced tier and a real-time tier is a practical way to serve two very different latency requirements without maintaining two separate research efforts.
- Migrating between model versions isn't automatic. Existing voice clones needed manual retraining to inherit v4's stability gains, which is a cost worth planning for before switching.
- Regenerating one piece of a larger work, a line or a paragraph, needs awareness of what surrounds it or the seam shows. That's a general lesson for any AI content workflow, not just audio.
- Time-boxed free access, 11 days here, is a low-friction way to get people to actually test a new model instead of just reading about it.
Terms worth knowing.
- Audio tags
- Bracketed delivery directions such as [whispers] or [laughs] inserted into a script to steer how a line is spoken.
- Instant Voice Clone
- A voice clone built from about 10 seconds of clean audio with immediate setup, used for quick, lower-fidelity results.
- Professional Voice Clone (PVC)
- ElevenLabs' higher-fidelity voice cloning option, most accurate when cloning a voice in its own native language.
- Context stitching
- The model's ability to account for the surrounding lines or chapter when regenerating a single line of narration.
- Bidirectional input streaming
- Sending text to the model as it's generated, for example by an LLM, while audio starts streaming back before the full text arrives.
- IPA (International Phonetic Alphabet)
- A phonetic notation used to specify exactly how a word should be pronounced, useful for names and tricky terms.
Things they pointed at.
Lines you could clip.
“Don't tell anyone, but we're breaking into the world's biggest candy store.”
“11v4 was designed to read a script the way a voice actor reads one.”
“If you hadn't noticed yet, this entire video is my 11v4 voice, made with an avatar entirely inside 11creative.”
Word for word.
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
The bait, then the rug-pull.
ElevenLabs opens with a direct claim: this is its most expressive text to speech model yet, and the on-camera host saying it is an AI avatar running on the new voice itself.
Named ideas worth stealing.
v4 vs v4 Turbo
- Eleven v4: produced content (narration, character work, voiceover, dubbing)
- Eleven v4 Turbo: real-time (voice agents, live assistants, interactive experiences)
Same underlying research split into two latency tiers so the model fits either a production pipeline or a live conversation.
How they asked for the click.
“Click the first link in the description, create a free account and pick 11v4. For the next 11 days you can use 11v4 for free.”
Direct, time-boxed free-trial CTA delivered by the same AI avatar host, paired with an on-screen 'Eleven days' title card and a description link.
Visual structure at a glance.
