The same sentence can comfort a patient, calm an angry customer, or rally a squad.
The words may be identical. The delivery is not.
That gap between what is said and how it lands is what ElevenLabs is targeting with Eleven v4 and Eleven v4 Turbo, launched today as its new flagship text-to-speech models.
ElevenLabs describes v4 as its most expressive speech model yet, built around a new architecture that interprets scripts more like performances than blocks of text. Turbo brings that same model family to real-time applications, where latency matters just as much as voice quality.
Both are available now through ElevenCreative, ElevenAgents, and the ElevenLabs API.
Voice AI is already fluent.
What it still struggles with is sounding convincingly aware of the moment.
A customer-support bot may deliver an apology with the same cadence it uses to confirm an order number. An audiobook voice can lose consistency across a long chapter. Dubbing can preserve the words while losing the personality of the original performance.
Eleven v4 is ElevenLabs’ attempt to solve those problems inside one model by combining expression, voice identity, context, and multilingual speech without forcing users to choose between them.
And with Turbo, ElevenLabs is making a second bet: that real-time speech no longer has to sound like the lower-quality version.
The early benchmarks are strong.
ElevenLabs says listeners preferred Eleven v4 roughly 75% of the time in blind head-to-head comparisons against Cartesia Sonic 3.6, Inworld TTS-2, and Google Gemini 3.8 Flash and Flash-Lite TTS.
At launch, v4 also topped Artificial Analysis’s Voice Arena with an Elo score of about 1,319, ahead of Sonic 3.6 at roughly 1,276 and Gemini 3.8 Flash TTS at around 1,267.
That leaderboard will move, so the important point is not that ElevenLabs will remain permanently at the top. It is that v4 is launching as one of the strongest-performing speech models currently available.
Eleven v4, identified in the API as eleven_v4, is the quality-first model.
ElevenLabs positions it as the beginning of a new model generation rather than a minor upgrade to v3.
The biggest shift is contextual delivery.
v4 is designed to infer tone, pacing, emotion, character, and surrounding context instead of simply turning words into polished speech.
A line like:
“I need you to stay calm.”
could sound reassuring, urgent, clinical, nervous, or sarcastic depending on what comes before it.
That makes the model especially interesting for storytelling, advertising, games, support conversations, and character-driven content where delivery carries as much meaning as the words themselves.
ElevenLabs has expanded its inline audio controls.
Creators can use directions such as:
[laughs]
[whispering]
[said angrily in French accent]
[light rain]
[phone buzzing]
Multiple directions can also be combined, and users can describe delivery in natural language.
ElevenLabs says v4 follows these instructions better than v3, although the company also acknowledges that tag adherence is still being improved.
At the same time, some familiar controls are disappearing.
The old Style and Speed sliders are gone, as is SSML support. The main controls are now Stability and Similarity, with more of the performance itself expected to come from the prompt and script.
Instant Voice Clones can now be created from roughly 10 seconds of audio, while Professional Voice Clones remain fully supported.
But v4 is also more sensitive to the quality of that source material.
Because the model tries to reproduce not just timbre but also cadence, loudness, vocal habits, and imperfections, a poor recording can carry more of those flaws into the final clone.
That means clean source audio matters more than ever.
It also means a v4 clone may sound noticeably different from a v3 clone created from the same person. In many cases, the v4 version should sound closer to the actual source recording rather than the interpretation produced by the older model.
Language coverage has grown from roughly 70 languages to 90+.
New and improved support includes languages such as Cantonese, Mongolian, and Odia, while ElevenLabs says some of the biggest quality gains appeared in Japanese, Brazilian Portuguese, Mandarin, and Cantonese.
That matters not only for accessibility but also for dubbing, localization, international customer support, and multilingual agents.
One of the more important changes is also one that may surprise existing users.
When the input language matches the voice’s native language, v4 attempts to preserve the original accent.
But when the model speaks a different language, it now prioritizes natural fluency in the target language rather than carrying the source accent into it.
For dubbing and international agents, that is likely an improvement.
But if you specifically wanted, say, English spoken with the accent of a Korean voice clone, the new behavior may not be what you want.
ElevenLabs says an accent-preservation toggle is still being researched.
v4 also increases the maximum generation length to 10,000 characters, up from 5,000 in v3.
Other improvements include better stitching across long-form requests, more natural multi-speaker dialogue, contextual reactions to previous lines, and improved IPA handling for difficult or custom pronunciations.
For audiobook creators, podcasters, game studios, and long-form narration teams, those improvements may matter as much as the headline gains in expressiveness.
If v4 is optimized for quality, Eleven v4 Turbo is optimized for immediacy.
Its API model ID is eleven_v4_turbo.
ElevenLabs reports median inference latency of about 100 milliseconds and median time to first speech of roughly 150 milliseconds using WebSocket streaming, excluding network latency.
For comparison, the company lists approximately 262 ms for Cartesia Sonic 3.6 and 814 ms for OpenAI GPT-4o mini TTS in the same benchmark.
The significance is straightforward.
A voice agent cannot feel conversational if the user finishes speaking and then waits for the system to visibly—or audibly—process the response.
At around 150 ms to first speech, ElevenLabs argues that Turbo can begin responding within the natural rhythm of human conversation.
Turbo also supports bidirectional streaming, meaning speech can begin before the underlying language model has finished generating the complete response.
That is especially useful for customer-service agents, games, assistants, and other applications where silence quickly becomes noticeable.
The goal is not just faster speech.
It is faster speech that still understands that a confirmation, apology, escalation, joke, or warning should not all sound the same.
In simple terms:
Use Eleven v4 when voice quality is part of the product.
Think audiobooks, advertisements, character performances, localization, narration, and premium creative work.
Use Eleven v4 Turbo when responsiveness is part of the product.
Think phone agents, interactive assistants, in-game characters, and real-time conversational systems.
ElevenLabs is launching the models with aggressive introductory pricing.
Standard API pricing is:
Using ElevenLabs’ rough estimate that 1,000 characters equals about one minute of generated speech, that works out to approximately $0.08 and $0.04 per minute respectively.
Through October 12, 2026, however, both models are discounted by 72%.
During the promotion:
That is equivalent to approximately $22 and $11 per million characters.
Creator-tier customers and above also receive access to v4 inside ElevenCreative at no additional cost, up to twice their monthly credit allocation.
UI generations still consume credits, with roughly one character corresponding to one credit before any applicable multipliers.
For production use, it is worth checking ElevenLabs’ live pricing page before building cost assumptions around the launch promotion.
The model IDs are:
eleven_v4
eleven_v4_turbo
A free ElevenLabs account is enough to begin experimenting.
If you already use Eleven v3, the best evaluation is not to start with a completely new demo. Start with material you already know well.
A practical test sequence:
ElevenLabs says v4 will continue training after launch, so outputs may change over time.
For anyone producing a large catalog of voice content, that makes periodic re-testing important.
The speech-generation market is moving quickly.
Cartesia, Inworld, Deepgram, Fish Audio, Google, OpenAI, and others are all pushing toward the same destination: voices that are more expressive, more controllable, multilingual, and fast enough for real-time interaction.
ElevenLabs’ advantage has increasingly been the breadth of its stack.
It is not only selling a TTS model. It has built voice cloning, creative tools, conversational agents, dubbing, and APIs around the same ecosystem. The company has also said that more than half of its revenue now comes from large enterprise customers.
That makes v4 more than a research demo.
It is a production upgrade.
There are new model IDs, new defaults, different accent behavior, updated pricing, and a limited-time discount clearly designed to encourage developers and creators to test the move away from v3.
The benchmark numbers will matter.
But the more interesting test is what happens outside the benchmark.
If Eleven v4 performs this well on real customer calls, audiobooks, games, ads, and multilingual content, the conversation around text-to-speech may shift again.
The question will no longer be whether AI-generated speech can sound human.
It will be whether your product is ready for a voice that understands the scene.
