ElevenLabs is an AI audio platform that turns written text into lifelike speech, clones real voices with permission, dubs video into other languages in the speaker’s own voice, and powers conversational voice agents that answer phones and chat on websites. If you only need one sentence to decide whether it matters to you, it is this: it is the tool most creators reach for when the work is a voice, not a chatbot, and when how something sounds matters as much as what it says.
ElevenLabs is an AI voice platform: software that generates human-sounding speech from text, reproduces a specific person’s voice from a recording, translates spoken content into other languages, and holds spoken conversations through AI agents.
Checked 5 November 2026. ElevenLabs changes its plans, models and features often, so treat the prices and model details here as current at that date and re-check the company’s own pages before you pay for anything. One honest caveat up front: a direct fetch of ElevenLabs’ own pricing page was blocked during the research for this guide, so the pricing figures below are the figures independent checks reported from that page on 1 November 2026, not figures read on the page today. They are consistent across those checks, but prices are the fastest-moving facts in this article, and the page itself is the final authority. No tool was hands-on tested for this article, so quality judgements are labelled as judgement, based on documented capabilities and the company’s own claims.
1. The short answer: what ElevenLabs is
ElevenLabs began as a text-to-speech company and grew into a full voice platform. That history still explains its shape. Most AI companies bolt a voice feature onto a chatbot or an assistant. ElevenLabs did the reverse: it built the voice first, then built everything around it — cloning, dubbing, agents, sound effects, music and transcription — on top of the same speech models. If you have heard an AI narrate a video this year, there is a good chance it was theirs, and our company explainer on what ElevenLabs is and how it got here sets out that rise in more detail.
The platform now splits into three surfaces, and keeping them straight saves confusion later. ElevenCreative is the content studio most readers will use: text-to-speech, the voice library, voice cloning, dubbing, music, sound effects and a long-form editor called Studio. ElevenAgents is the conversational side: a platform for building voice and chat agents that handle real customer conversations on the phone, on a website, or over WhatsApp. ElevenAPI is the developer route to the same capabilities, for teams that want to embed voices into their own apps rather than work in ElevenLabs’ dashboard. One account and one credit pool sit underneath all three, which is why the pricing section later in this guide matters to every kind of user, not just creators.
The scale is worth stating plainly, because it answers the question of whether this is a toy or infrastructure. ElevenLabs crossed $500 million in annual recurring revenue in May 2026, and investors valued the company at $22 billion in September 2026, figures covered in our explainer on how ElevenLabs makes money from dubbing, agents and audiobooks. Those numbers do not make its voices good — only listening does that — but they do mean the platform is not a weekend project. It is a business with enterprise customers, media deals and a developer ecosystem, and its roadmap moves accordingly: fewer novelty features, more reliability, compliance and control.
Who is it actually for? Four groups, roughly. Creators who narrate videos, podcasts and courses. Publishers and studios turning text catalogues into audiobooks or dubbed versions. Businesses that want a voice agent answering routine calls. And developers building voice into products, from reader apps to games. This guide walks through each group’s tools in turn, then the pricing that ties them together, then the consent rules that govern all of it. By the end you should be able to answer the only question that matters: which part of ElevenLabs, if any, fits the job you actually have.
<img src="
” alt=”A single studio microphone in front of a soft glowing pattern of abstract sound waves”>
2. How text-to-speech actually works
Text-to-speech, or TTS, converts written words into spoken audio. The old version of the technology stitched together fragments of recorded speech, which is why it sounded like a lift announcing floors. The modern version is a neural network trained on many hours of recorded speech. It does not just pronounce words; it predicts how a person would perform them — where to pause, what to emphasise, how a sentence should rise and fall. That performance layer is why today’s AI voices sound human where older systems sounded mechanical, and it is the whole reason ElevenLabs exists. For the general technology behind this, see our explainer on AI voices, cloning and text-to-speech, which separates the two core technologies and where each is used.
The workflow in ElevenCreative is deliberately simple, and that simplicity is part of the product. You paste or type your script, choose a voice from the library, choose a model, and generate audio. Then you listen, adjust, and download. The adjustments are where the craft lives. Voice settings typically include stability, which controls how consistent the voice sounds across a passage; similarity, which controls how closely the output hews to the chosen voice; and style, which controls how much the model exaggerates the speaking style. Higher stability sounds more even but can turn flat on unusual words. Lower stability sounds more natural but can wander. There is no universally right setting; narration, advertising and character work want different values, and the honest way to find yours is to generate the same paragraph at two or three settings and listen.
Models are the second choice, and they trade expressiveness against speed and cost. The most expressive model in the current line is Eleven v3, reported as covering 70 or more languages and supporting audio tags — written directions such as [whispers] or [sarcastically] placed inside the script to steer delivery. That tag system matters more than it first appears: it turns the script itself into the direction, so a writer can shape a performance without touching a slider. At the other end sits the Flash family, built for speed at roughly 75 milliseconds of latency, with fewer languages (around 32 reported) and lower cost per character, aimed at real-time agents and bulk jobs where waiting is worse than a slightly flatter read. Between them sit conversational variants tuned for dialogue. The practical rule is simple: use the expressive model when the audio is the product, and the fast model when the audio is a means to an end, such as an agent answering a question.
One more mechanism deserves a plain explanation, because it explains both the quality and the limits. The model generates the whole performance at once rather than assembling recorded fragments, which is why breath, rhythm and intonation survive. But it is still predicting likely speech, not consulting a verified source. It can mispronounce an unusual name, flatten a joke, or rush a sentence that needs a pause. Punctuation helps: commas, full stops, exclamation marks and ellipses all steer pacing, and the documentation notes that descriptive wording can influence expression. The craft of using ElevenLabs well is mostly the craft of writing scripts that direct well — short sentences for emphasis, punctuation that marks breaths, and tags where the model supports them — then listening like an editor, not accepting the first take.
3. The voice library: picking a voice that fits
Before you clone anything or write a script, you will probably spend time in the voice library, and it deserves a section of its own because it is where most users actually start. The library is reported to hold more than 10,000 voices — premade voices plus thousands of community-shared voices — searchable by language, accent, gender, age and use case. That scale is both the appeal and the trap. A library that large contains the right voice for almost any job, but finding it takes filtering, not browsing. Search by language and accent first, then by use case, then listen to the same test sentence across your shortlist rather than judging voices on different scripts.
Two kinds of voice live in the library, and the distinction matters for rights. Stock and premade voices are designed or licensed characters you can use under your plan’s terms. Community voices are shared by other users, and their availability and terms depend on how they were shared. A third route, Voice Design, lets you describe a voice in words — age, accent, tone, character — and have the system generate one to match. That is useful when you need a voice that does not exist yet, such as a specific character for a series, and it avoids the rights questions that come with cloning a real person, because no real person is involved.
Judgement, not documentation, decides most voice choices, so here is the judgement that holds up across use cases. For narration, pick the voice you can listen to for an hour without fatigue, not the one that impresses in ten seconds; warmth and steadiness beat drama. For advertising, pick the voice that matches the brand’s energy in the first three seconds, because no one hears the rest if the first three feel wrong. For characters, pick distinctiveness over realism — a voice that is recognisably different from your other characters matters more than one that sounds generically human. And for anything multilingual, check the voice in each target language, not just your own: a voice that sounds natural in English can sound strained in another language, and the documentation itself recommends matching the voice’s accent to the target language and region rather than treating every voice as interchangeable.
A final practical point: voices are not portable files. A voice you design, clone or favourite lives inside your ElevenLabs account, and you use it through the platform or its API. That is fine for most work, but it means switching providers later means re-casting, not exporting. If a voice becomes the sound of your show, that lock-in is worth knowing about before you publish a hundred episodes in it.
4. Voice cloning: instant vs professional
Voice cloning creates a digital model of a specific person’s voice from a recording of them speaking, so that the model can then read any new text in that voice. This is the most powerful feature in ElevenLabs and the most sensitive, so it gets one careful section here — and deliberately not a full walkthrough. That walkthrough is a separate queued guide, and this product map stays at the level of what cloning is, which kind to choose, and what the rules are.
There are two kinds, and they are quite different products. Instant Voice Cloning builds a usable clone quickly from a short sample — around a minute of audio, with the documentation warning against exceeding about three minutes for this route. It is fast, it is available from the Starter plan upward, and for short-form narration and prototyping it is often good enough. Professional Voice Cloning is the serious version: it trains a dedicated, higher-fidelity model from 30 minutes or more of clean, high-quality recorded audio, captures accent, emotional range and vocal character more precisely, and is available from the Creator plan upward. The trade is time and effort against fidelity. If the clone only needs to read short scripts convincingly, instant is enough. If it needs to carry an audiobook, a character, or a brand voice across hours of material, professional is the honest choice.
Quality depends overwhelmingly on the sample, and the rules for a good one are unglamorous. One speaker only, no background noise, no room echo, consistent volume and tone, natural speaking rather than performed announcing. The documentation’s advice is consistent across every guide that summarises it: speak normally, at a normal pace, with normal inflection, because the clone is learning you, not a character. File format matters less than cleanliness, though lossless formats are preferred where you have them. A noisy three-minute sample will produce a worse clone than a clean one-minute sample, and no setting rescues a bad recording.
Then the rule that governs everything else in this section. Professional Voice Cloning is for your own voice. It uses a voice-captcha step — a phrase spoken live — to verify that the uploaded audio matches the person creating the clone. You cannot use this pathway to clone a third party’s voice, even with their consent, according to the documentation as summarised in the guides checked for this article. Instant cloning carries its own consent requirements, and misusing cloning violates ElevenLabs’ terms and can lead to a permanent ban. The full consent and safety picture is in section 13, and the practical cloning walkthrough, with recording and quality checklists, is coming as its own guide. For this map, the point is simple: cloning is a tool for scaling your own voice, not for borrowing someone else’s.
5. Dubbing: the same voice in another language
Dubbing is ElevenLabs’ answer to a problem every creator with an international audience knows: your video works, but only in one language. Traditional dubbing hires new actors per language, re-records everything, and takes weeks. ElevenLabs’ Dubbing Studio ingests a video or audio file, transcribes it, translates the script, and generates dubbed audio in the original speaker’s voice — preserving, as far as the technology allows, the timing, tone and performance of the original. The result is a version of your video in another language that still sounds like you, not like a stranger reading your lines.
The workflow follows the same pattern as the rest of the platform: upload a video or audio file, or paste a video link, choose target languages, and let the system detect speakers, translate and re-voice. Output options reported for dubbing include video files, audio files and subtitle files, so the dubbed version can go back into an editor for finishing. It works best on clean, single-speaker content; overlapping speech, heavy background music and noisy recordings all make the transcription and translation stages harder, and the dub inherits those difficulties. As with cloning, the quality of the input sets the ceiling for the output.
Language counts need an honest note, because sources disagree and the disagreement has a reason. Classic dubbing coverage is reported at 29 or more languages with speaker-voice preservation, while an August 2026 changelog for the project-based dubbing API describes translation into more than 90 languages. Both can be true: different dubbing routes, at different dates, cover different language sets. The same caution applies to the headline language figures elsewhere in this guide. When a specific language matters to your project, check it in the product itself before you plan a launch around it, rather than trusting any single number in any single guide, including this one.
Pricing for dubbing is also different from text-to-speech, and it catches people out. Dubbing is billed per source minute — the length of the original, not the output — and the per-minute cost varies by plan and by the dubbing route. Free-tier dubs carry a watermark. The practical consequence is that dubbing a long back catalogue is a budgeting exercise before it is a creative one: price the catalogue first, test one representative video end to end, and only then commit to the batch. A dubbed video that sounds right but costs more than the revenue it earns in that language is a bad dub, however good the voice sounds.
6. Voice agents: ElevenAgents, the phone-answering side
Everything so far has been about generating audio files. Voice agents are different: they hold conversations. ElevenAgents is the platform for building them — voice and chat agents that answer customer support calls, qualify sales leads, book appointments, track orders and handle routine enquiries, across phone lines, website widgets, WhatsApp and contact-centre platforms. Where text-to-speech reads a script, an agent listens, thinks, and replies, turn by turn, in real time.
An agent is assembled from parts, and understanding the parts explains both the power and the failure modes. The voice is the sound of the agent. The language model is its brain, deciding what to say. The knowledge base is what it is allowed to know — your help articles, policies, product details. Tools are what it can do — check an order, book a slot, escalate to a human. Guardrails are what it must not do — the boundaries that keep it from promising refunds you do not offer or discussing topics outside its job. Change any one part and the agent behaves differently, which is why agent work is operations work, not a one-time setup: prompts, knowledge and guardrails need maintenance as your business changes.
Deployment channels are the practical question most buyers get wrong. A phone agent and a website widget agent are not the same product with a different skin. Phone calls have interruptions, background noise, accents, and callers who talk over the agent; the turn-taking problem — knowing whether a caller has finished speaking or is only saying “mm-hmm” — is the difference between an agent that feels natural and one that feels robotic. Web agents have cleaner audio but impatient users who expect instant answers. WhatsApp agents live inside a messaging app with its own etiquette. ElevenAgents covers all of these, but the testing regime for each is different, and a team that only tests on the website will be surprised by the phone.
The competitive context matters here too. Voice agents are where the whole voice industry is consolidating: model makers, infrastructure companies and open-source frameworks are all racing to own more of the stack. Our report on Deepgram’s acquisition of Daily and the Pipecat framework shows the rival shape of the same bet — an open orchestration layer plus a real-time network — and it is a useful comparison when you evaluate ElevenAgents, because it frames the choice as closed-and-integrated versus open-and-assembled. Neither is right for everyone; the question is which failure mode you would rather manage.
7. Eleven Realtime: when the agent listens and talks in one model
Until recently, building a voice agent meant stitching three separate systems together: one to transcribe what the caller said, one to decide what to reply, and one to speak the reply aloud. Each handoff added latency and a chance for the conversation to feel disjointed. Eleven Realtime, launched on 20 October 2026, folds listening, reasoning and speaking into a single speech-to-speech model, and it is the most technically significant launch in this guide. Our launch report on Eleven Realtime’s Mini and Max models covers the details, and this section summarises what matters for a buyer.
There are two models. Mini is the cheaper, conversation-first model, built for natural, expressive talk — assistants, tutors, interactive characters. Max adds reasoning for harder work: customer support, lead qualification and multi-step workflows that rely on chained tool calls, and it carries the compliance features — Zero Retention Mode and HIPAA support — that Mini does not. The company says realtime conversation works in more than 90 languages, drawing on its voice library, and that timing and tone stay natural through interruptions and quick back-and-forth exchanges. Those are the company’s claims, reported by the outlets that covered the launch; no independent verification of the benchmark figures had been published at the time of that report, and the same caution applies here.
There are two ways in, and they do not mix. Inside ElevenAgents, a builder switches an agent’s voice mode to Realtime and chooses Mini or Max, keeping the existing prompt, voice, tools and knowledge base. Through the Realtime API, a developer points an existing OpenAI Realtime client at ElevenLabs by changing the endpoint and model name, with client and server events in the same shape. That compatibility layer is a deliberate bid for switchers: a team already built on OpenAI’s realtime API can trial ElevenLabs without rebuilding. The limits matter for planning: the Realtime API connects over WebSocket only at launch, agents using a custom language model or certain workflow nodes must stay in the older mode, and the models are rolling out gradually, so some workspaces will not see the option yet.
Related to agents, and easy to confuse with Realtime, is ElevenAgents Architect, released in alpha on 19 October 2026. Architect is not a voice model; it is a conversational builder that lets teams create and repair agents using plain-language requests instead of dashboard menus, analysing conversation logs to find failing workflows and drafting fixes that a human reviews before they go live. Our report on ElevenAgents Architect in alpha explains the draft-and-publish safeguard, which is the detail that matters most: nothing reaches live callers until a person publishes the draft. If you are evaluating ElevenAgents, Architect is the feature that most changes the operating cost, because agent maintenance, not agent creation, is where enterprise deployments usually stall.
8. Speech to text, Scribe and the Cue timing layer
Voice work runs in both directions, and ElevenLabs’ speech-to-text side is easy to overlook because the company is famous for the opposite direction. Scribe is its transcription service: feed it audio and it returns text, across many languages, including a real-time variant for streaming. Transcription is the unglamorous foundation of several other features — dubbing starts with a transcript, agents need to hear callers accurately, and searchable audio archives depend on it — so its quality sets a ceiling for everything built on top of it.
Sitting behind transcription is a quieter piece of engineering that deserves a mention because it changes what transcripts can do. Cue is a forced-alignment model: given a transcript and a recording, it works out exactly when each word — and, with Cue, each character — was spoken. That timing layer is what makes transcripts clickable, subtitles appear on cue, and dubbed dialogue land in sync with the speaker on screen. Our report on ElevenLabs’ Cue and character-level audio timing covers its benchmark claims, including an average timing error the company reports at 179 milliseconds and coverage of more than 90 languages in a single model. As with all vendor benchmarks, treat those as the company’s figures, not independent verdicts, but the practical point stands: timing turns a transcript from a document into an interface.
For a creator, the workflow value is concrete. A captioned video, a searchable podcast archive, a dubbed video whose subtitles actually match the new audio, and a transcript a listener can click to jump to the exact moment — all of these depend on alignment, not just transcription. For a developer, the same layer is available through a forced-alignment API, although the company had not stated pricing for that access in the sources checked. The honest summary is that speech-to-text is not ElevenLabs’ headline product, but it is the connective tissue of the platform, and a team that only evaluates its text-to-speech is evaluating half the system.
9. Sound effects and music: finishing the soundtrack
A voice alone rarely finishes a video. Sound effects and music are the rest of the soundtrack, and ElevenLabs has moved into both. Sound effects are generated from text prompts: describe the sound — a door closing, rain on a window, a crowd cheering — and the system generates it, priced per generation from the same credit pool. For creators, that replaces searching stock libraries for a near-enough effect; for the platform, it extends the same generative-audio technology from speech into the rest of the soundscape.
Music goes further. ElevenMusic generates original music, and Instant Soundtracks, introduced in October 2026, scores an uploaded video with original music and timed sound effects in a single pass: it analyses the footage scene by scene, writes a score to fit the picture, and places effects at the matching timestamps, with effects on separate layers from the music so an editor can adjust them without starting over. Our report on ElevenLabs’ Instant Soundtracks sets out the API limits that matter for planning — up to ten videos, 200 MB combined, and ten minutes of video per request — and the licensing caution that belongs with every generated track: a track made for a personal test is not automatically cleared for advertising, so check what your account includes before using output in paid media.
The pairing with video matters beyond ElevenLabs itself. AI video tools increasingly generate silent footage that still needs a soundtrack, and our comparison of AI video generators from Runway, Sora, Veo, Kling and Pika shows how often sound is the missing half of a generated clip. Judgement: ElevenLabs’ sound effects are a genuine time-saver for short-form work, and Instant Soundtracks is promising for creators publishing daily, but a generated score can be in sync yet wrong for the story. Editors should still judge the result rather than accepting it blindly, because timing, taste and brand fit are hard to automate.
10. Studio: long-form projects, audiobooks and courses
Short clips are one job. Audiobooks, courses, training libraries and long-form narration are another, and they break tools designed for one paragraph at a time. Studio is ElevenLabs’ long-form editor: a project workspace where a whole book, course or script lives as one project, broken into chapters or sections, with voices assigned per speaker or character, and edits made at the level of a sentence rather than by regenerating everything. The plan tables reflect that role: the Free plan includes 3 Studio projects, Starter raises that to 20, and higher plans raise it further.
The workflow difference is worth spelling out, because it is where long-form projects succeed or fail. In a single-clip tool, a correction means regenerating the clip. In Studio, a correction means editing one paragraph and regenerating that paragraph, keeping the rest of the performance intact. Character consistency across a long project — the same voice, the same pacing, the same pronunciation of names — is the hard problem, and it is solved less by settings than by discipline: fix pronunciations early, keep a list of names and how they should sound, and listen to chapter boundaries, where a voice can drift. The tools that treat narration as a performance, with sentence-level editing and consistent voices, are the tools professionals stay with, and Studio is ElevenLabs’ version of that discipline.
Publishing is the other half of the long-form story. Audiobook producers, news sites and education companies convert text catalogues into narrated audio at a fraction of studio cost, and for publishers sitting on decades of un-narrated backlist, that is not a novelty; it is a new revenue line on assets they already own. That is also why the commercial licence boundary in the pricing section matters so much to this group: a backlist narrated on a Free plan cannot be sold, and the plan that unlocks commercial use is the first real decision a publisher makes, before any question of voice or model.
11. The API: putting ElevenLabs voices inside your own product
Some readers will never open the ElevenLabs dashboard after the first week. They are developers, and for them the product is the API: the same text-to-speech, cloning, dubbing, transcription and agent capabilities, accessible programmatically so they can be embedded into apps, games, reader tools, marketing software and support systems. The core endpoint is a REST call — a POST request to the text-to-speech endpoint with a voice identifier — authenticated with an API key, with official Python and JavaScript or TypeScript software development kits wrapping it. If you have built against any REST API before, the shape will feel familiar; the learning curve is in voice behaviour, not in HTTP.
API pricing works differently from app credits, and confusing the two is the most common budgeting mistake in this whole guide. In the app, text-to-speech costs credits from a monthly pool, at one credit per character on standard models. Through the API, usage is priced in US dollars per thousand characters, reported at $0.08 per thousand on the standard models and $0.04 on the Turbo and conversational variants, with speech-to-text reported at $0.22 per audio hour for the standard transcription model. The two meters describe the same underlying work in different units, and the API pricing page is the place to reconcile them for your volume. Estimate API volume separately from app volume, because a product that generates speech for thousands of users does not behave like a creator generating a few hours a month.
The strategic point about the API is stickiness. Once a voice is wired into a product — a reader app’s default narrator, a game’s character voices, a support tool’s agent — it tends to stay, because changing it means re-casting and re-testing every interaction. That is why the API matters to ElevenLabs’ business beyond its direct revenue: it seeds the other streams, as a startup that starts with a few thousand characters of API usage can grow into a media or enterprise contract. For a developer evaluating it, the questions to ask are the boring ones: latency at your volume, output formats your player supports, concurrency limits on your plan, and what happens to your users if a model version changes. The API reference answers those; marketing pages do not.
12. Pricing and credits, explained honestly
Pricing is where most guides about ElevenLabs go wrong, usually by quoting a single number as if it were the whole story. The real system has two meters, six self-serve plans, an enterprise tier, and a set of first-month offers designed to be misread. This section gives the figures as reported from ElevenLabs’ own pricing page on 1 November 2026, with the caveat from the top of this guide: a direct fetch of that page was blocked during research, so re-check the page itself before you pay. Prices exclude taxes.
The Creative subscription ladder, monthly billing, is as follows. Free costs $0 and includes 10,000 credits a month, roughly 10 minutes of text-to-speech, with no commercial licence. Starter costs $6 a month and includes 30,000 credits, roughly 30 minutes, and is the first plan with a commercial licence, Instant Voice Cloning, 20 Studio projects, Dubbing Studio and music commercial use. Creator costs $22 a month — shown on the card at $11 for the first month only — and includes 121,000 credits, roughly 121 minutes, adding Professional Voice Cloning. Pro costs $99 a month and includes 600,000 credits, roughly 600 minutes, adding 44.1 kHz PCM output through the API and 192 kbps audio quality. Scale costs $299 a month for 1,800,000 credits and 3 seats, and Business costs $990 a month for 6,000,000 credits and 10 seats. Enterprise is priced on request. Annual billing charges ten months for twelve, giving monthly equivalents of $5, $18.33, $82.50, $249.17 and $825 across Starter to Business.
Three traps deserve emphasis because they cost real money. First, the $11 Creator figure is month one only; month two is $22, and budgeting $11 forever is how a surprise invoice happens. Second, credits are a shared pool, not a per-product allowance: the same balance pays for text-to-speech, transcription, dubbing, music and sound effects, at different rates per product, so a month of heavy dubbing can drain a balance a text-to-speech user would have found generous. Dubbing is billed per source minute, music at a per-minute credit rate, sound effects per generation, and transcription per minute — check the current per-product rates on the pricing page, because they move. Third, unused credits on paid plans are reported to roll over for up to two months, but not on Free, so a Free account that waits for a big month will find its balance has not grown.
So which plan should you actually buy? The honest decision rule is volume plus licence, not features. If you are testing, Free is enough to learn the workflow, but nothing on it can be used commercially. If you publish anything for money — videos, audiobooks, client work, ads — Starter is the floor, because it is the first plan with a commercial licence. If you need a professional clone of your own voice, Creator is the floor, because that is where Professional Voice Cloning unlocks. If you produce at volume or need the highest audio quality and API output options, Pro is the floor. Everything above Pro is a team or enterprise decision about seats and volume, not about features a solo creator needs. And one more practical note: agents and API usage are priced on separate tabs of the pricing page from the Creative subscriptions, so a team buying agents should price that surface separately rather than assuming the Creative ladder covers it. Agent per-minute pricing is not published as a single figure in the sources checked, so this guide does not estimate one.
13. Consent, safety and ethics: the rules that keep voices legitimate
Every powerful feature in this guide has a misuse case, and voice cloning has the sharpest one: a faithful copy of a voice is, to a listener’s ear, the person. That is why the consent rules are not a footnote to ElevenLabs but the condition that keeps its market legitimate, and why this guide treats them as a full section rather than a disclaimer.
The rules, as documented and as summarised in the guides checked, are straightforward. Professional Voice Cloning is for your own voice, verified by a live voice-captcha step. Cloning another person without permission violates ElevenLabs’ terms and can lead to a permanent ban. Instant cloning carries its own consent requirements, and the platform operates an AI Speech Classifier to help detect audio generated by its own system. Generated audio belongs to the user who generated it, but commercial use and monetisation rights are restricted to paid subscriptions — the licence boundary drawn in the pricing section. None of this is optional decoration; it is the operating licence for the whole product.
Beyond the platform’s rules sit the law and the audience’s expectations, and both are moving in the same direction: disclosure. The EU AI Act’s transparency rules, in force since August 2026, require AI-generated content — including synthetic voices — to be identifiable as machine-made, a point our general AI voices explainer covers in its section on the rules catching up. Other jurisdictions are writing disclosure rules of their own. The practical advice that follows from all of it is simple. If a voice is selling, informing or persuading, assume listeners deserve to know when it is synthetic. Label it. Do not use a cloned voice to imply an endorsement the person never gave. Do not use a voice agent that pretends to be human when a caller asks directly. And keep records of consent for every real person’s voice you use, because the question will not be whether you had a good reason; it will be whether you can show permission.
The industry’s argument about voice actors deserves a fair hearing here, because it is the human side of the same question. Voice actors are competing with replicas of voices — sometimes, with consent and payment, their own. The legitimate version of this market pays voice owners, licences library voices properly, and labels synthetic audio. The illegitimate version clones without asking and pockets the difference. ElevenLabs’ position in that argument is helped by its consent verification and detection tools, and hurt every time those tools are bypassed or ignored. For a buyer, the three checks that separate serious platforms from liabilities are the ones our voices explainer names: rights (are library voices licensed, and is consent required before cloning a real person?), control (can you direct performance and edit at sentence level?), and disclosure (does the platform make it easy to label synthetic audio and detect its own voices?). A service that will clone any voice from any clip without questions is a liability dressed as a feature, whatever its price.
14. Frequently asked questions
What does ElevenLabs text-to-speech do? It converts written text into spoken audio in a chosen voice, with control over pacing, emphasis and emotion, so a script becomes narration, a voiceover or an audiobook chapter without a recording session.
What does ElevenLabs voice cloning do? It builds a reusable model of a specific person’s voice from a recording, so that voice can read any new text. Instant cloning works from a short sample; professional cloning uses 30 minutes or more of clean audio for higher fidelity, and is verified as your own voice by a live voice captcha.
What does ElevenLabs dubbing do? It translates a video or audio programme into other languages and re-voices it in the original speaker’s voice, preserving timing and tone as far as the technology allows, with subtitles and audio outputs for finishing in an editor.
What do ElevenLabs voice agents do? They hold spoken conversations — answering support calls, booking appointments, qualifying leads — by combining a voice, a language model, a knowledge base, tools and guardrails, deployed on phone lines, websites, WhatsApp or contact-centre platforms. Realtime models let one model listen and speak instead of stitching three systems together.
Can I use ElevenLabs for free? Yes, for testing and personal, non-commercial use: the Free plan includes 10,000 credits a month. Commercial use requires a paid plan, starting with Starter at $6 a month as reported on 1 November 2026 — re-check the pricing page before buying.
Can I clone someone else’s voice? Not through the professional cloning route, which verifies your own voice by captcha, and not without permission under ElevenLabs’ terms. Misuse risks a permanent ban, and the ethical and legal exposure sits with the person who made the clone, not just the platform.
How many languages does ElevenLabs support? It depends on the product: the expressive text-to-speech model is reported at 70 or more languages, the fast model at around 32, classic dubbing at 29 or more, and the newer dubbing API and realtime agents at 90 or more. Check your specific language in the product before planning a launch around it.
Is ElevenLabs better than the big labs’ voice tools? It depends on the job. For dedicated voice work — cloning, dubbing, long-form narration, voice agents — its specialisation, voice library and controls are the argument. For a voice feature inside an assistant you already pay for, the bundled tools from OpenAI, Google or others may be enough. Our explainer on OpenAI, from GPT-3 to GPT-5, gives the context for how the big labs’ voice capabilities arrived inside their assistants.
The verdict is refreshingly simple for a platform this large. ElevenLabs is the best-documented, most complete voice toolkit available to an ordinary creator or a small team in 2026, and the right way to use it is to buy the smallest plan that gives you the licence and the feature you actually need, test one real project end to end, and only then scale. Start with the voice library before you clone anything. Dub one video before you dub a catalogue. Run one agent on one phone line before you automate a call centre. The voices are good enough that the limiting factor is rarely the model; it is the script, the sample, the consent, and the discipline to listen to the result like an editor. Get those four right, and ElevenLabs will sound less like a tool and more like the voice your work was missing.
0 Comments