ElevenLabs Launches Eleven v4 and v4 Turbo: Voice Models That Can “Perform the Story” from Text

ElevenLabs has released two new voice models: Eleven v4 and Eleven v4 Turbo. Eleven v4 targets finished content such as audiobooks, character voices, animation, advertising, and podcasts, with an emphasis on emotion, pacing, character performance, and long-form consistency. v4 Turbo targets customer-service calls and voice agents, retaining those expressive capabilities while reducing time to first audio to a level suited to real-time conversation.

Most AI voices today no longer sound obviously bad. What still breaks the illusion is something else: they know how to pronounce every word, but not why a line should be delivered in a particular way.

A character has just been frightened, yet the voice remains flat. A customer-service agent has found good news, but sounds exactly as it did while asking the caller to keep waiting. Re-generate only the third line of a scene, and the character may suddenly sound like a different actor. Eleven v4 targets that gap: moving text to speech beyond “reading correctly” toward “performing from the script.” v4 Turbo is meant to bring the same expressiveness into live calls and voice agents in time for a natural exchange.

Key features of Eleven v4

• The script can directly control the performance. Emotion, pauses, laughter, whispers, actions, and ambient sound effects can be written into the dialogue, alongside natural-language directions for how a line should be spoken.
• Voice cloning balances similarity and expressiveness. Instant Voice Clone can create a voice from roughly 10 seconds of audio, while Professional Voice Clone again supports v4's full emotional range.
• Long-form generation and re-generation are more coherent. Context stitching connects pacing and state across generated segments, while speaker stability aims to preserve the same voice across narration, dialogue, and replacement lines.
• Multi-speaker dialogue and pronunciation control. One script can coordinate several speakers, turn-taking, and sound effects, while IPA and pronunciation dictionaries lock down names, acronyms, and specialist terms.
• More than 90 languages. ElevenLabs specifically names Cantonese, Mongolian, and Odia; its demonstrations also show Catalan narration with sound-effect tags.
• Turbo cuts the wait for real-time conversation. ElevenLabs reports median model inference of about 100 ms and median time to first audio of about 150 ms, with bidirectional streaming support.

Performance control in the script: emotion, pauses, actions, and sound effects

Traditional TTS scripts usually contain little beyond the words to be spoken. To make a voice tenser, gentler, or better paced, creators either adjust model parameters or generate many takes and select one. v4 moves part of the director's job into the text itself: a script can include [whispers], [laughs], [long pause], and [door slams], or a natural-language direction such as “deliver this line as if you have just discovered the truth.”

These tags can appear in sequence to build an entire scene. In the quiz-show example on the product page, the host begins warmly, gradually builds tension, sounds a gong, and leaves five seconds of silence. The contestant quietly guesses in a British accent with uncertainty and nervous laughter. Once the answer is confirmed, the host, applause, and teasing follow naturally. The script no longer specifies only what is said; it also says who is speaking, what just happened, and how the line should land.

The horror monologue further down the launch thread makes the point more clearly. Its on-screen script includes [exhaling], [chuckling], and [smacking lips]. The same voice first threatens the listener at close range and counts their breaths while they sleep, then shifts into a choked, affectionate recollection of its mother. The challenge is not merely inserting “laughter,” but changing emotion with the character relationship rather than firing a series of sound-effect buttons.

There is one practical migration detail: in v4, pauses are mainly controlled with natural-language audio tags such as [pause] and [long pause]. The product-page FAQ explicitly says that SSML tags such as <break> are not supported in v4. Teams with large SSML script libraries cannot finish the migration by changing only the model ID; they also need to rewrite pause and performance instructions.

Voice cloning: preserve the voice while allowing emotion to change

Voice cloning has always involved competing goals: sound like the original person while retaining enough emotional range; allow heightened performance without losing the underlying vocal identity.

ElevenLabs states the contrast with earlier models directly in its demonstration: v2 was relatively similar to the speaker and stable; v3 was much more expressive and supported audio tags and advanced prompting, but the presenter felt it “doesn't quite sound like me.” v4's goal is to combine the two, preserving the same person's identity while emotion changes within a sentence.

Instant Voice Clone in v4 requires only about 10 seconds of audio. Professional Voice Clone, absent from v3, is supported again and can use v4's full emotional range. But the product page also contains an easy-to-miss condition: IVC or PVC voices created before v4 launched must be retrained to work effectively with v4. For teams with a large library of branded voices, this is not an entirely frictionless replacement.

This comparison demonstrates v4's product goal: a cloned voice should still sound like the same person while laughing, whispering, gasping, or changing emotion mid-sentence. In practice, the quality, noise level, and speaking state of a 10-second recording will affect the result, so teams should still test with their own samples.

Long-form continuity: maintaining pacing and character identity across segments

Long-form audio is rarely generated in one pass. An audiobook is produced chapter by chapter and segment by segment, and a misread line may need to be regenerated on its own. Each generation, however, is a new start: speed, pitch, and emotion can drift, and a replacement sentence may sound as though another actor has suddenly taken over.

v4 addresses this with two capabilities. The first is context stitching, which carries pacing and performance state across neighboring segments so separate generations sound like one recording session. The second is speaker stability, intended to preserve a speaker's identity across dialogue, narration, and regenerated lines. The product page describes support for scripts of any length, while the launch thread uses the phrase “infinite text generation.”

In practice, a single generation still has a 10,000-character limit. “Any length” does not mean placing unlimited text in one request; it means repeatedly segmenting the script and relying on contextual stitching to preserve continuity. For audiobooks and long-form narration, the key workflow change is that segmented generation and single-line repairs should no longer require rebuilding the character state from scratch.

v4 versus v4 Turbo: finished content and real-time conversation

An advertising voice-over can wait a few extra seconds while the best take is selected, so quality and performance come first. If an AI falls silent for a second after the caller finishes speaking, the phone experience is already damaged. The distinction between v4 and Turbo is fundamentally about where the compute budget is spent.

The two latency figures should not be collapsed into one. About 100 ms refers to the median inference time the model needs to begin generating audio; about 150 ms is the median interval from request to the arrival of the first audio segment. A complete AI call also involves speech recognition, LLM reasoning, network round trips, and playback buffering, so end-to-end latency will necessarily exceed the TTS model's figure alone.

Turbo's key companion is bidirectional streaming: the upstream LLM need not wait for a complete sentence before sending text to TTS, and audio can be returned as it is generated. What this reduces is not the time needed to finish the entire recording, but how soon the user hears the first word. The minimal JavaScript example in the launch thread sets modelId to eleven_v4_turbo, but that image shows a standard textToSpeech.convert() call rather than the full bidirectional streaming interface.

The main official video places Turbo in a pharmaceutical authorization call. The agent must correctly handle the physician's name and a specialist term such as semaglutide; after confirming that the prescription has been approved, it should ask to switch to Mandarin to explain the medication to the caller's mother. The scene combines pronunciation, character stability, conversational timing, and language switching in one call, making it closer to an actual voice-agent workflow than a standalone narration sample.

Multi-speaker dialogue, IPA, and pronunciation dictionaries

v4 supports several speakers and sound effects in a single script. In the baseball commentary demonstration, two broadcasters joke between plays: one reveals that the other once choreographed for the Backstreet Boys, and the other makes fun of his own age and “rizz.” A moment later the game resumes, both return to high-energy home-run commentary, and crowd cheering fills the background.

A multi-speaker script is more than alternating voices. Each person must retain an individual identity while sharing the current situation, knowing what the previous speaker said, where the joke landed, and when play resumed. This directly corresponds to the product page's claim that the model knows who is speaking, what just happened, and how each line should be delivered.

Pronunciation is controlled through IPA and pronunciation dictionaries. Names, acronyms, and technical terms can be assigned a reading in advance and reused across generations. The baseball clip in the launch thread demonstrates multi-speaker interaction; IPA support comes from the feature descriptions on the same thread and the product page.

More than 90 languages: the same performance controls for non-English scripts

Both models are claimed to support more than 90 languages, with the launch thread specifically naming Cantonese, Mongolian, and Odia. A companion clip narrates in Catalan and includes [música de transició] on screen, showing that non-English input can combine performance and sound-effect directions.

For teams with production requirements in Mandarin, Cantonese, or less widely supported languages, the language count is only the first layer. Accent, prosody, specialist pronunciation, and cloning similarity still determine whether a result can ship, so blind listening tests with real voices and production scripts remain essential.

Artificial Analysis leaderboard: currently ranked first

The opening post says Eleven v4 ranks first on Artificial Analysis. When we read the live Provider Voice Arena leaderboard, it did rank first with an Elo of 1315, a 95% confidence interval of ±18, 1,693 comparison samples, and 8 voices. Cartesia Sonic 3.6 followed at 1275 and Gemini 3.8 Flash TTS at 1267.

This is not a vendor-assigned score. Listeners hear two clips generated from the same text without seeing model identities, choose which sounds more natural, and the Elo score updates from those pairwise preferences. It therefore supports a specific conclusion: within the Provider Voice Arena's current blind-listening sample, Eleven v4 ranks first in human preference.

The ranking measures overall naturalness preference in blind listening. It does not directly test cloning similarity, Chinese scripts, long-form drift, real telephone networks, or cost. Elo also changes as new votes arrive: the public snapshot on launch day was 1319, while our reading had already moved to 1315. The result makes v4 a strong candidate to test first, but each production scenario still needs its own acceptance testing.

Five checks before production use

First, old scripts are not necessarily drop-in compatible. SSML <break> tags must be replaced by v4's natural-language pause tags, and older IVC/PVC voices may need to be retrained.

Second, long-form does not mean unlimited input in a single request. Each generation is capped at 10,000 characters; longer content still needs segmentation, with context stitching trying to carry state across the boundaries. The most important test is not how impressive the first segment sounds, but whether chapter ten still sounds like the same person and whether a replacement line can be inserted seamlessly.

Third, real-time performance must be measured across the entire chain. Turbo's roughly 100 ms and 150 ms figures both describe the TTS layer. A voice agent must also hear the caller, understand the request, invoke tools, and buffer playback. Customer-service testing should cover interruptions, long numbers, names, drug names, background noise, and cross-language switching.

Fourth, pricing must be read with its time limit. During the two-week launch promotion, ElevenLabs listed discounted API pricing of $22 per million characters for v4 and $11 for Turbo. Creator+ users could use v4 free in ElevenCreative up to a limit equal to twice their monthly credits. This was a limited launch offer, not a permanent price to place in a long-term cost model. The standard product page also listed 10,000 monthly credits on the free tier and said that this was roughly enough for 10 minutes of audio.

Fifth, cloned voices raise consent and retention questions, not only quality questions. ElevenLabs says both instant and professional cloning require verification by the voice owner, and that scripts and audio tags are not used for training without the user's consent. Eligible enterprise services can enable Zero Retention Mode, but only when the specific service supports it and the option is properly enabled; “enterprise account” should not be treated as automatically meaning that no data is retained.

Eleven v4 writes part of the voice director's job into the script

What makes Eleven v4 most notable is not simply another realistic-sounding TTS model. It attempts to compress some of the intent previously distributed among actor, director, sound design, and post-production back into a script that can be written directly: who the character is, what just happened, whether this line should be restrained or explosive, where to pause, when to laugh, and what is happening in the background.

v4 turns those directions into deliverable content; Turbo says them before the conversation goes cold. Both are available across ElevenCreative, ElevenAgents, and ElevenAPI, spanning character performance through real-time agents.

The official demonstrations show emotion, multiple speakers, sound effects, cloning, and language switching in one generation workflow. Artificial Analysis's blind-listening leaderboard, meanwhile, indicates strong human preference for v4's overall naturalness.

The most useful next test is to give the model your own difficult passages: one character across chapters, an emotional reversal within a sentence, two people talking over a scene, Chinese proper nouns, or a single line regenerated and inserted back into the whole. v4's value lies not only in the first impressive listen, but in whether the character retains the same voice and performance state after repeated revisions.