What is Eleven v4?
Eleven v4 is the highest-quality text to speech model from ElevenLabs, released on September 28, 2026. It reads text in one voice or performs a full multi-speaker dialogue, follows inline audio tags like [whispering] and [laughing], and speaks 70+ languages. On UnificAlly it runs as elevenlabs/eleven-v4 and is billed per character.
It is the first model in a new ElevenLabs generation and the successor to Eleven v3. The jump shows on the public board: on the Artificial Analysis Text to Speech arena, Eleven v4 scores an Elo of 1328, against 1176 for Eleven v3.
What's new in Eleven v4
Top two on the Artificial Analysis TTS arena
In blind listening votes, Eleven v4 scores an Elo of 1328 and Eleven v4 Turbo 1330, the top two places on the board, ahead of Qwen-Audio-3.1-TTS-Plus (1300), Sonic 3.6 (1286) and Gemini 3.8 Flash TTS (1274), as of October 14, 2026.
Closer voice clones
Pass a few audio samples of one speaker as voice and v4 clones the voice for that request only, then deletes it. It picks up timbre, cadence and delivery more faithfully than v3, so a clone sounds like the person, not a rough match.
Audio tags for acting
Write directions like [quietly curious], [annoyed] or [soft chuckle] in the text and the voice performs them in place. Narration, game lines and character scenes get direction without a separate script.
Dialogue in one request
Send a dialogue list of lines, each with its own voice ID, and get back one audio file for the whole scene, with the turn-taking handled by the model.
10,000 characters per request. Twice the limit of Eleven v3, enough for a long scene or a book section in one call.
Native accents across languages. When you generate in a language that differs from the voice's own, v4 speaks the target language fluently instead of carrying the original accent over.
Captions included. One-voice requests can return character alignment and word timings next to the audio.
What Eleven v4 is best for
Audiobooks
Long narration with tagged pacing and tone, in chunks of up to 10,000 characters.
Game and animation voices
Character lines with emotional range, directed through audio tags.
Dubbing and localization
One voice across 70+ languages, speaking each one without a foreign accent.
Branded voice clones
A voice built from a few clean samples, used for one request and then removed.
Podcasts and explainers
Two or more hosts rendered as one dialogue file.
Eleven v4 use cases
Narrate an audiobook chapter by chapter, with tags like [measured] or [lower, thoughtful] setting the pace of each scene. Voice every quest-giver in a game from one model, with a tagged line for each mood. Clone a presenter from a short, clean recording and voice their course in several languages without the accent drifting. Turn a two-host script into a finished podcast segment with one dialogue request, then use word timings from a narration request to cut captions.
Eleven v4 limitations
Audio tags work well but are not followed every time yet. ElevenLabs is still improving tag adherence. There are no style or speed controls and no SSML, so pacing comes from the text and the tags. Clones copy their source closely, flaws included, so noisy samples give noisy voices. The cross-language accent behavior is always on: if a voice must keep its native accent in another language, test it first. ElevenLabs also keeps training v4 after launch, so output can shift over time.
Eleven v4 vs Eleven v4 Turbo
Both take the same input and sit within two Elo points of each other on Artificial Analysis (1328 vs 1330, overlapping ranges). Use Eleven v4 Turbo when response time matters, such as voice agents and live characters, or when a lower per-character rate matters more than the last bit of polish. Use Eleven v4 for narration, dubbing and finished voiceovers where quality comes first.
Eleven v4 API examples
curl -X POST https://api.unifically.com/v1/tasks \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "elevenlabs/eleven-v4",
"input": {
"text": "[whispering] The door was already open... Someone had been here.",
"voice": "21m00Tcm4TlvDq8ikWAM",
"with_timestamps": true
}
}'
Poll GET /v1/tasks/{task_id} until status is completed, then read output.audio_url. Dialogue, voice cloning and every output format are covered at docs.unifically.com.
FAQs
People also ask about Eleven v4
Send a POST /v1/tasks request with model set to elevenlabs/eleven-v4. Put text and a voice ID in input for one voice, or a dialogue list of text and voice pairs for several speakers. You get a task_id back, and the finished task returns an audio_url.
ElevenLabs released Eleven v4 and Eleven v4 Turbo on September 28, 2026. Eleven v4 is the higher-quality model of the pair, and Turbo is the real-time one.
On the Artificial Analysis Text to Speech arena, Eleven v4 sits 2nd with an Elo of 1328, two points behind Eleven v4 Turbo at 1330 and ahead of Qwen-Audio-3.1-TTS-Plus at 1300 (as of October 14, 2026). Eleven v3 sits at 1176 on the same board.
Yes. Pass a list of audio sample URLs of one speaker as voice and Eleven v4 clones that voice for the request only, then deletes it. Up to 25 samples are accepted. Clean recordings matter, because v4 copies the source closely, including room noise and mic quirks.
Each request takes up to 10,000 characters of text. Longer scripts, such as a full audiobook, need to be split into chapters or scenes.
Only stability and similarity_boost. Lower stability gives a more varied, expressive read, and higher similarity keeps the output closer to the reference voice. Style, speed and speaker boost do not exist on v4, and SSML is not supported.
Eleven v4 is billed per character of text you send, at $0.048/1000 characters. A new account starts with a small free balance, so you can try it before you top up.
Yes. Set with_timestamps to true on a one-voice request and the result carries character alignment plus a word list with start and end times, ready for captions or lip-sync. Dialogue requests return the audio only.
