What Is ElevenLabs? AI Voice Generation

ElevenLabs is a voice AI company that turns text into speech realistic enough that listeners often cannot tell it from a person, and that can clone a voice from a short recording. It was founded in 2022 and is now a platform with several product lines: speech synthesis, voice cloning, transcription, music, and phone-ready voice agents. This article explains what the company sells, how that kind of technology works, what it costs as of October 2026, who uses it, and why the same realism makes it a fraud and consent problem.

The company in numbers

ElevenLabs was started in 2022; Mati Staniszewski is named as a co-founder in press coverage of the Series D.

The money story moved quickly. On February 4, 2026, the company announced a $500 million Series D at an $11 billion valuation, led by Sequoia Capital, with a16z and ICONIQ also participating. Business Today reported total funding of $781 million across five rounds and described the valuation as roughly triple the figure a year earlier. ElevenLabs says it ended 2025 with more than $330 million in annual recurring revenue.

Those figures are company-reported or press-reported rather than audited. They are useful mainly as a sense of scale: this is no longer a hobbyist text-to-speech website.

What ElevenLabs sells

ElevenLabs groups its products into three lines, per its Series D announcement.

• ElevenCreative for generating, editing, and localizing audio: speech, music, sound effects, voice cloning, and dubbing. The company says it supports more than 70 languages.
• ElevenAgents for building voice and chat agents that handle customer operations over phone, chat, email, and WhatsApp.
• ElevenAPI for developers who want low-latency voice infrastructure inside their own products.

Beneath those sit specific models. The company's model documentation lists the following lineup as of this writing.

Model	Job	Languages	Notes from the docs
Eleven v3	Text to speech	70+	5,000-character limit
Multilingual v2	Text to speech	29	10,000-character limit
Flash v2.5	Text to speech	32	About 75 ms latency, 50% lower API price per character; 40,000-character limit
Eleven v4	Text to speech	90+	10,000-character limit
Eleven v4 Turbo	Text to speech, low latency	90+	About 100 ms median inference latency
Scribe v2	Speech to text	90+	Speaker labels for up to 32 speakers
Scribe v2 Realtime	Live speech to text	90+	About 150 ms latency

The "75 ms" figure that circulates in marketing refers to the Flash model and, as the documentation notes, excludes application and network delay. Your real-world latency will be higher.

Model names change quickly, and older models such as Turbo v2.5 are already marked deprecated. Treat any model list, including this one, as a snapshot.

How AI voice generation works

Older text-to-speech systems were concatenative: engineers recorded a voice actor reading thousands of sentences, sliced the recordings into sound fragments, and stitched fragments together at run time. The result was clear but robotic, because stitching cannot capture how a sentence's emotional shape changes every sound inside it. Think of a ransom note assembled from magazine letters.

Modern systems are closer to a language model for sound. One influential design, published in January 2023 as VALL-E, is called a neural codec language model. It compresses audio into a sequence of discrete codes using a neural audio codec, then trains a language model to predict the next code the way a text model predicts the next word. VALL-E was trained on about 60,000 hours of English speech and could imitate an unseen speaker from a three-second recording used as a prompt.

ElevenLabs has not published the full architecture of its production models, so the following is the general shape of the field, not a description of its internals. Most current systems do three things.

1. Read the text. Text is split into tokens, small chunks that the model processes; our explainer on how tokenization works covers the idea.
2. Condition on a voice. A short reference clip is turned into a numerical summary of the speaker's timbre and style, often called a speaker embedding.
3. Generate and decode. The model predicts audio codes or spectrogram frames, and a decoder (a vocoder) turns them into a waveform you can play.

The same scaling pattern seen in text applies: more data and larger models generally improve naturalness. For the mechanics of that learning process, see how neural networks work and how large language models work. A useful comparison is image generation: just as diffusion models replaced hand-built image filters, learned audio models replaced hand-built speech synthesizers.

Why "expressive" models need tags

Plain text carries little information about delivery. "Fine." can be agreement, anger, or resignation. Some recent expressive models accept inline audio tags, short instructions in brackets that steer emotion or timing. This is a trade: you gain control, but you also have to direct the performance, much as a film director gives notes to an actor rather than only handing over a script.

Voice cloning: two tiers

Cloning is the feature that draws the most attention, and the most worry. ElevenLabs offers two tiers, with details from its voice cloning documentation.

	Instant Voice Cloning	Professional Voice Cloning
Audio needed	Roughly 1 to 2 minutes of clean audio recommended	30 to 180 minutes recommended
Training time	Near-instant	Typically 3 to 6 hours, longer depending on queue
Verification	None mentioned in the docs	Required to confirm the voice is yours
Plan needed	Starter and above	Creator and above

The documentation states that "even with their consent, you cannot clone someone else's voice" with the professional tier, because the verification step requires the speaker to prove ownership. Instant cloning has no comparable check described in the docs, which is the gap critics focus on.

The current documentation recommends one to two minutes of good audio for instant cloning, so claims of much shorter samples are best read as best-case figures. Why does more audio help? A clone has to capture more than pitch. It has to learn how the speaker's vowels shift, how fast they talk, where they breathe, and how they stress words in questions versus statements. A one-minute clip of calm reading shows few of those habits, so a clone built from it can sound right on neutral sentences and strange on shouting or laughter. Longer, varied recordings give the system examples of the range. This is the same reason a caricaturist needs more than one photo to draw a convincing likeness in motion.

What it costs (October 2026)

The pricing page lists these plans on monthly billing. Prices and credit amounts change often, so confirm on the pricing page before budgeting.

Plan	Monthly price	Credits per month	Cloning
Free	$0	10,000	None listed
Starter	$6	30,000	Instant
Creator	$11 first month (50% off)	121,000	Professional (1 voice)
Pro	$99	600,000	Professional (1 voice)
Scale	$299	1,800,000	Professional (3 voices)
Business	$990	6,000,000	Professional (10 voices)
Enterprise	Custom	Custom	Custom

A credit is a unit of usage, and how many credits a clip consumes depends on the model and the feature, so the monthly totals do not translate into a fixed number of minutes. The free tier does not list a commercial license; commercial use begins at Starter, according to the pricing page.

Subscription AI pricing drifts, with quotas and prices shifting while plan names stay the same, which is a reason to treat any dated price table as a snapshot, as our piece on AI chatbot pricing also illustrates.

To budget, check the credit rate for your chosen model on the pricing page and multiply by your script length. Retakes are the hidden cost: because generation is probabilistic, the same sentence can come out slightly different each time, and polishing a script often means regenerating lines.

Who uses it

The February 2026 announcement named customers by product line:

• ElevenAgents: Deutsche Telekom, Square, the Ukrainian government, and Revolut.
• ElevenCreative: Duolingo, NVIDIA, and TIME.
• ElevenAPI: Meta, Epic Games, Salesforce, MasterClass, and Harvey.

These are company-announced relationships; the scope of each deployment is not given.

Typical uses

• Narration and audiobooks. Authors and publishers produce spoken versions without booking studio time. Quality on long passages depends on the model, and per-request character limits mean long scripts must be split.
• Localization. A creator's voice can be rendered in other languages. For tools aimed at translating text, see our overview of AI translation tools.
• Games and apps. Studios can generate placeholder or final dialogue. The labor politics of this are covered below.
• Customer service. A voice agent answers phone calls and takes simple actions. These systems combine speech recognition, a language model, and speech synthesis in one loop, which is why latency matters; AI agents explained covers the agent layer.
• Accessibility. People who lose speech can use a recorded sample of their own voice. This is among the most widely cited humane uses of voice cloning in general, though results vary by sample quality.
• Transcription. The Scribe models handle the reverse direction, speech to text. If transcription is your need, compare options in best AI transcription tools, and see why transcription tools make different mistakes.

How a voice agent answers a phone call

The agent product is where latency becomes the central engineering problem, so it deserves its own walk-through. A phone agent is usually a relay race with four runners.

1. Listening. A speech-to-text model converts the caller's audio into words as they speak. ElevenLabs lists a realtime Scribe model with about 150 ms latency for live transcription.
2. Deciding when the caller has finished. The system must detect the end of a turn. Cut in too early and the agent interrupts; wait too long and the pause feels awkward.
3. Thinking. A language model reads the transcript, the conversation so far, and any business rules, then drafts a reply or calls a tool, such as looking up an order.
4. Speaking. A text-to-speech model turns the reply into audio, ideally starting to play before the full sentence is generated.

Each step has its own delay, and they add up. Human conversation tends to have gaps of a fraction of a second between turns, so an agent that takes two full seconds to respond feels broken even if every answer is correct. That is why vendors advertise tens of milliseconds for the speech step: it is one slice of a budget the whole chain must share. A fast voice cannot compensate for a slow language model, which is why many deployments use smaller, quicker models for simple questions and escalate harder ones, a pattern described in our explainer on model routing.

The business case follows from that structure. A voice agent can answer unlimited simultaneous calls at the marginal cost of computing, but it can also confidently give a wrong answer with a friendly voice. Guardrails, test suites, and human hand-off matter more than voice quality once the audio is good enough.

Where synthetic speech still fails

Even strong models stumble on predictable material, and knowing the failure types helps you check output efficiently.

• Numbers, dates, and units. "1/2" can be a fraction or a date. "Dr." can be doctor or drive. Spell these out when accuracy matters.
• Names and loanwords. Uncommon surnames, brands, and place names are mispronounced more often than common words, because the model has fewer examples. Pronunciation dictionaries and respelling fix most cases.
• Homographs. "Read" in the present versus the past tense depends on context the model may misjudge.
• Long-form drift. Over many minutes, pacing and timbre can wander. Splitting long scripts into shorter requests, within each model's character limit, is a common workaround.
• Emotional mismatch. Text rarely states the intended emotion, so the model guesses. Audio tags in the newer expressive models exist to correct this.
• Cross-language cloning. A voice cloned from English may carry an English accent into another language, because the clone captures the speaker's habits as well as their pitch.

A good editing routine is to listen once at normal speed for flow, then re-read the script against the audio for names and numbers, since those are the errors listeners notice and repeat.

The competitive landscape

ElevenLabs competes on several fronts at once, which is why a single scoreboard is misleading.

Competitor type	What they compete on	Examples to compare
Frontier AI labs	Bundled voice with their chatbots, often lower price	OpenAI, Google
Low-latency specialists	Speed for live agents	Cartesia
Creator platforms	Editing and podcast workflows	Descript, Speechify
Music models	Song generation	Suno

Any numeric ranking, such as a score from expert blind tests, should be treated with suspicion unless the test design is published, including who listened, how many clips, and which text. For a current, hands-on comparison across vendors, see our best AI voice tools ranking, and for a broader content stack, AI tools for content creation.

Risks and controversies

Fraud and political deepfakes

The best-known abuse case is the January 2024 robocall in New Hampshire that imitated President Biden's voice and urged Democrats not to vote in the primary. In May 2024 the FCC proposed a $6 million fine against the political consultant behind it, and a $2 million fine against the carrier that transmitted the calls, Lingo Telecom, according to Nextgov. The case shows what cheap synthetic voices make possible, whichever vendor's tools are involved.

The lesson is structural. Any system that lowers the cost of a convincing voice also lowers the cost of impersonation. A realistic voice in a phone call is like a forged signature on a letter: the forgery gets cheaper to make, so the recipient's habits of verification have to get stronger.

Detection and provenance

ElevenLabs offers an AI Speech Classifier. Its page says the tool "reliably identifies audio generated with most ElevenLabs models," does not reliably classify audio from the Eleven v3 model, cannot detect audio from other providers, and analyzes only the first minute of each sample. In other words, the tool can say "probably ours," not "definitely human."

The classifier page does not mention an audio watermark, so listeners should not assume one is present in a given clip.

Consent and the entertainment industry

Voice actors have organized against unconsented digital replicas. The SAG-AFTRA video game strike ended after members ratified a new Interactive Media Agreement on July 9, 2025 with 95.04% support, according to SAN's report. The contract requires written consent before studios create digital replicas and includes disclosure and compensation terms. The strike itself had begun in July 2024.

ElevenLabs has pursued a licensing route for famous voices. In November 2025 it announced an Iconic Marketplace in which companies can request approval to use voices of performers and estates, with Michael Caine among those announced. Variety reported that companies request approval to use a listed voice and, once both sides agree, formalize a license; separately, Matthew McConaughey is investing in the company and producing a Spanish audio edition of his newsletter in his own voice. That is consent by licensing, which differs from the instant cloning of an ordinary user's own recording.

What people get wrong about AI voices

"It needs hours of audio." Not for a usable clone. Instant cloning works from a minute or two, but the professional tier, which asks for 30 minutes or more, is intended for higher fidelity.

"Detectors will solve this." Classifiers have clear limits: they are tuned to one vendor's output, degrade with post-processing, and run in a losing race against new models. The practical defense for individuals is out-of-band verification, such as calling back on a known number or agreeing a family code word.

"The voice is the person." A clone reproduces timbre and some mannerisms. It does not reproduce the person's knowledge or intent, which is why a convincing voice on the phone is weak evidence of identity.

"Cheap voices kill voice acting." The evidence is mixed and contested. Labor agreements are now writing consent and compensation rules into contracts, which suggests the market is adapting rather than collapsing, but the long-term effect on voice work is not settled.

"All AI voice is real-time." Generating a minute of narration and holding a live phone conversation are different problems. The second needs speech recognition, a language model, and speech synthesis to finish within a few hundred milliseconds. For why that pipeline is hard, see AI inference versus training.

How to evaluate a voice AI tool

If you are choosing a tool, a short checklist works better than a leaderboard.

1. Listen to your own text. Demos are tuned. Paste a paragraph from your own material, including names and numbers, and listen for mispronunciations and odd pacing.
2. Check license terms for your use. On ElevenLabs, commercial use starts at the Starter plan. Read the terms before publishing.
3. Know the limits. Per-request character limits differ by model, from 5,000 characters for Eleven v3 to 40,000 for Flash v2.5, so long scripts must be split.
4. Understand consent rules. Clone only your own voice or a voice you have written permission to use.
5. Plan for drift. Models and prices change often, so build your workflow so you can swap providers.

Studies of preference between synthetic and human voices are difficult to generalize, because results depend on the text, the listener, and the clip length. Short clips are far easier to fake than long ones, where cadence and breathing patterns drift.

Frequently Asked Questions

Q: What is ElevenLabs used for? A: It is used to generate speech from text, clone voices, dub and localize content, transcribe audio, generate music, and build voice agents. Publishers, game and app developers, and customer-service teams are typical users.

Q: Is ElevenLabs free? A: There is a Free plan with 10,000 credits per month. Commercial use and instant voice cloning begin on the $6 per month Starter plan, as of the pricing page reviewed for this update.

Q: How much audio does voice cloning need? A: Instant cloning works from about one to two minutes of clean audio. Professional cloning recommends 30 to 180 minutes and requires verification that the voice is yours.

Q: Who founded ElevenLabs? A: It was founded in 2022; Mati Staniszewski is named as a co-founder in press coverage of its February 2026 funding round.

Q: How is ElevenLabs valued? A: In February 2026 it raised $500 million at an $11 billion valuation, led by Sequoia Capital, according to the company's announcement and press coverage. Valuations change with new rounds, so check recent reporting.

Q: Can AI-generated audio be detected? A: Partly. ElevenLabs offers a classifier that reliably identifies audio from most of its own models, but it cannot detect audio from other providers.

Q: Does ElevenLabs support real-time conversation? A: Yes. Its Flash model lists about 75 ms latency and its v4 Turbo model about 100 ms median inference latency, both excluding network and application delay.

The takeaway

ElevenLabs shows how quickly a single technique, predicting audio the way language models predict text, turned speech synthesis from a novelty into a commercial platform with an $11 billion valuation. The same capability produces both the accessibility tool and the impersonation tool, and the safeguards (verification for professional clones, classifiers, licensing contracts) are partial. For a reader, the practical conclusions are modest: treat unverified voices on phone calls with the skepticism once reserved for unverified emails, and treat any dated claim about pricing, models, or customers as a snapshot.

Sources

• ElevenLabs, Series D announcement (February 4, 2026)
• Business Today, ElevenLabs bags $500 million in funding (February 5, 2026)
• ElevenLabs documentation: models, voice cloning, and pricing
• Wang et al., Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers (January 2023)
• Nextgov, Mastermind behind Biden AI robocalls faces potential $6M fine from FCC (May 2024)
• SAN, Video game actors end 11-month strike with new AI protections (July 2025)
• Variety, Matthew McConaughey, Michael Caine ink deals for AI-generated versions of their voices (November 2025)