ElevenLabs review: voice production and editing

Editor's note: This is a documentation-led edition. We have not yet run our own hands-on tests of ElevenLabs. What follows rests on official documentation, published pricing and independent studies we read in full, all accessed on 16 October 2026. A tested update will follow.

This review looks at ElevenLabs as a production tool: where a producer imports a script, picks or clones a voice, edits narration chapter by chapter, dubs it and exports a master. Our profile of ElevenLabs and the AI voice business covers the company itself. Here the question is narrower: what can you make, at what cost, and under which rules?

We make no claims about how the voices sound. But the documentation sets out what a producer needs before a first session: hard limits, plan gates, licence carve-outs and data terms. If you are setting up a process rather than choosing a tool, start with our guide to building an AI voiceover workflow with review.

What it is, and what each plan is for

ElevenLabs sells a web app, ElevenCreative, and an API. Producers mostly work in the app.

The current models

Eleven v4, released on 28 September 2026, is now the flagship, with a real-time Turbo variant.[1] The main text-to-speech models:

Source: ElevenLabs models page.[2] Latency is the company's figure, excluding network time.

Two v4 changes matter for production. Style and Speed sliders are gone, and "SSML is not supported." And a cloned voice no longer carries its native accent into other languages, which the company calls "a deliberate change in how cross-language generation works, not a bug."[3]

Plans and credits

Every ElevenCreative plan draws on one monthly credit pool shared across products.[4]

Sources: pricing page, Studio and Zero Retention Mode docs.[4][5][6] First-month discounts shown that day are excluded.

Text to speech costs about 1 credit per character. Music costs 900 credits a minute, a sound effect 200 per generation, and dubbing 2,000 to 10,000 per source minute, depending on mode and watermark.[4] Unused credits roll over for up to two months on paid plans, but downgrading forfeits them.[4] One request can't use more than twice your monthly quota.[7] API use is billed in dollars: $0.08 per 1,000 characters for v4, and $0.04 for v4 Turbo and Flash.[8]

Studio: long-form editing

Studio 4.0 is the long-form editor. Audio projects are for "long-form narration, podcasts, and audiobooks", with chapters detected on import. Video projects add a timeline, captions and Studio Agent, an AI co-editor whose chat is billed in credits on top of what it generates.[5] For narration, the documented tools are:

• Takes and locking. Each paragraph keeps a generation history you can restore from; deleting a take is permanent. Two retakes are free if text, voice and settings are unchanged, and locked paragraphs can't be edited by accident.[5]
• Actor Mode. You record or upload a reading to guide delivery. The documentation warns it "will replicate all aspects of the audio you provide, including the accent."[5]
• Pronunciation dictionaries, and timestamped comments from reviewers who don't need an account.[5]
• Checks on bulk export. The company says export scans for distortion, voice drift, mispronunciations and missing or extra words, and regenerates up to twice at no cost.[5]

Studio has hard ceilings: 500 chapters per project, 400 paragraphs per chapter and 5,000 characters per paragraph.[5] Exports come as MP3 or WAV, whole or by chapter, with optional normalisation and ISBN metadata. Free, Starter and Creator plans get 128 kbps MP3, or WAV converted from a 128 kbps source.[5] The documentation also sets expectations about repeatability:

"It's important to note that the AI is non-deterministic; setting the sliders to specific values won't guarantee the same results every time."

— ElevenLabs documentation, ElevenCreative Studio.[5]

Voices: cloning tiers, consent and the library

ElevenLabs offers two kinds of clone, with different checks.

Instant Voice Clone (IVC), from Starter. You upload about one to two minutes of clean audio and confirm "that you have the right and consent to clone the voice."[9] That is a declaration, not a voice check. In March 2025, Consumer Reports assessed six cloning products and found ElevenLabs among four that "required only that researchers check a box confirming that they had the legal right to clone the voice".[10]

Professional Voice Clone (PVC), from Creator. This needs at least 30 minutes of speech, ideally two to three hours, and usually takes three to six hours to fine-tune. It can't sing, and it must be your own voice: a failed verification means a 24-hour wait or a support request.[11] ElevenLabs describes that step, Voice CAPTCHA, as comparing your samples with "a real-time recording of the user reading a dynamically generated script."[12] A narrator who wants a producer to use their voice must create the clone in their own account and share it.[11]

For long projects, note that a clone can't be exported, and remaking it from the same audio gives "slightly different" results.[9] Drop below Creator and your PVC can't be used until you upgrade.[11]

The Voice Library and payouts. Only PVCs can be published to the library, after a review by ElevenLabs staff.[13] Owners set a notice period of up to two years, which can be lengthened but never shortened. They can switch on Live Moderation, which screens scripts against prohibited categories.[13][14] For producers, the notice period is the continuity risk:

"If the owner of a voice without a notice period decides to stop sharing their voice, you'll lose access to the voice immediately."

— ElevenLabs documentation, Voice Library.[13]

Owners are paid through Stripe Connect, available in about 60 countries, roughly weekly once earnings pass a threshold, usually $10.[15] Rewards depend on the user's plan, characters generated and the notice period. Free-plan use earns nothing, agent use is capped at $0.01 a minute, and estimates are "not a guarantee of earnings."[14] Non-US owners face 30% US withholding on US earnings unless they claim a treaty rate.[15]

Dubbing, sound effects and music

Dubbing. Automatic Dubbing runs on the Dubbing v2 Alpha model across 90+ languages, with a 0–10 cloning-strength setting that trades resemblance against natural delivery.[16] Dubbing Studio, the per-clip editor, works only with the older v1 model and is "in maintenance mode"; v2 editing through the API needs Enterprise.[16][17] For broadcast delivery:

• Lip sync is not offered in Dubbing.[17]
• Audio is mono on v1 and at most stereo on v2, so 5.1 mixes are not preserved.[17]
• Video comes back mostly at 1080p. The FAQ says: "Dubbed video is a high-quality preview, not a production master."[17]

Sound effects run up to 30 seconds per generation, with looping, and WAV at 48 kHz for non-looping effects.[18] Eleven Music makes tracks of 3 seconds to 5 minutes, with or without vocals, and can regenerate single sections; paid plans can add a 30-second reference track, screened for copyright.[19]

What each plan lets you publish

The terms of use give free-plan users non-commercial use only; paid users may use the service commercially, subject to the prohibited-use policy.[20] Three products add their own limits:

• Dubbing v2. Without written authorisation, output can't be used in feature films, TV series, streaming or video-on-demand programming, or theatrical releases. Creator uploads to YouTube, TikTok or Instagram, and paid ads on streaming platforms, are allowed.[21]
• Music. On every self-serve plan, the rights table permits commercial use "except film, TV, radio, and Studio Games". Starter, Creator and Pro are "For Individual Use Only"; Scale is for organisations under 10 employees and Business under 50. Distribution to music-streaming services starts on Creator.[22]
• Sound effects may only be used combined with other material. They can't be sold or shared as standalone files or sample libraries.[21]

AI Chronicle

Safeguards a producer will meet

Our company profile covers ElevenLabs' safety layers. Three details bear on production.

No-Go Voices. Uploaded samples are screened against a list designed to block the voices of public officials, candidates, celebrities "and other widely recognized figures".[12][23] Blocked voices "are no longer usable."[23]

Watermarking and detection. ElevenLabs says it has begun working with Google DeepMind to bring SynthID watermarking to audio generated on its platform.[12] A free Audio Detector, for signed-in users, checks for the watermark and then falls back to the older AI Speech Classifier. It finds only ElevenLabs watermarks, and "content generated by third-party models available through ElevenLabs is generally not watermarked."[24] The company puts the classifier at 99% precision and 80% recall on unmodified audio, but says it "does not reliably classify audio generated with the Eleven v3 model."[24][25] Neither page mentions v4. Detection is one signal among several, as our guide to spotting AI-generated content explains.

Provenance. The safety page says ElevenLabs incorporates C2PA standards.[26] We found no documentation describing Content Credentials attached to Studio or API exports. For what signed provenance can and can't show, see our explainer on content credentials.

Data retention and training terms

Your rights and theirs. The terms say "you retain all rights in and to your Output", but grant ElevenLabs a "perpetual and irrevocable" licence to use your content, voice included, to improve its services and develop new products. It "will not commercialize your voice on a standalone basis without your permission."[20]

Training opt-out. The "Data use" menu lets you opt out, but only from then on.[20][27] Whether training on licensed audio is lawful is a separate question, which our explainer on AI training and copyright covers.

How long data is kept. The privacy policy says voice data is kept no more than three years after your last interaction, unless the law requires or allows longer.[27] Deleting a generation through the API removes it at once, though debugging and moderation logs may keep related data, and backups last up to 30 days.[6]

Zero Retention Mode deletes most request data after each call, but only for Enterprise customers on eligible API endpoints. "Traffic through the web UI is never covered," and Studio, dubbing and cloning are not eligible.[6] Some Studio features also run on third-party AI providers, whose terms apply.[21]

Consent records. If you upload an identifiable person's voice, you must keep proof of consent for five years, or longer if the law requires.[21]

Documented limits and risks

• Default models lag the flagship. Studio defaults new projects to Multilingual v2, and old projects keep the quality setting chosen when they were created.[5]
• Pronunciation rules may be skipped. The API guide says phoneme rules work only on v4, v3 and Flash v2; "Other models skip dictionary phoneme tags", which includes Studio's default, Multilingual v2. The Studio page names Flash v2 alone. Break tags max out at 3 seconds, and "some newer models may reduce or ignore" them.[28][5]
• Lock-in. Clones can't be exported, and library voices can disappear when an owner withdraws them.[9][13]
• Dubbing is not a broadcast master, for the reasons above.[17][21]
• Documentation conflicts. The Free plan shows 3 Studio projects on the pricing page but 5 in the Studio FAQ. The dubbing upload cap is 1 GB on one page and 2 GB on another.[4][5][16][17]
• Instant-clone consent is self-declared.[9][10]

Who it suits, and who should skip it

These judgements follow from the documentation, not from our own use.

It suits:

• Audiobook and course producers working in chapters, who need take history, locking and a reviewer link. Budget for Pro if you need 44.1 kHz WAV.
• Narrators licensing their own voice, in a Stripe Connect country, who accept that rates aren't guaranteed.
• Creators localising for YouTube, TikTok or ads, where Dubbing v2 is expressly allowed.

Skip it, or negotiate first, if:

• You dub for film, TV or streaming programming.
• Your scripts are confidential and you work in the app, where zero retention never applies.
• You need a portable voice asset.
• You need music for broadcast or a studio game on a self-serve plan.

How to test it yourself

1. Set data use first. Opt out of training in Data use before uploading anything unreleased.
2. Run one chapter on two models. Generate the same 2,000-character passage on Multilingual v2 and v4, logging credits charged and retakes needed.
3. Stress the dictionary. Add five names as alias rules and five as phoneme rules, and check which the chosen model honours.
4. Check your export. Confirm the file format, sample rate and bitrate your plan delivers, and measure loudness against your delivery spec.
5. Dub a short clip and check the output's channels and resolution before promising a client a deliverable.
6. Read the library terms for any library voice you plan to use, including its notice period, before a long project begins.

A worked example (hypothetical)

This scenario is invented. No tool was run, and the numbers are planning estimates, not measurements.

Tomasz, an independent producer, is narrating a 70,000-word non-fiction book for its author.

Voice. The author creates a Professional Voice Clone on her own Creator account and shares it with Tomasz by link, since he can't clone her voice himself.

Plan. At roughly six characters per word, including spaces, the text is about 420,000 characters. At 1 credit a character, that exceeds Creator's 121,000 credits but fits Pro's 600,000, leaving room for paid retakes. Pro also gives him 44.1 kHz WAV exports.

Editing. He imports the manuscript into a Studio audio project, splits any paragraph over 5,000 characters, and locks paragraphs as the author approves each chapter via a shared link.

Paperwork. He files the author's written consent, to keep for five years.

Questions readers commonly ask

Can I use ElevenLabs for paid work on the free plan?

No. The terms limit free users to non-commercial use. Commercial use starts on Starter, at $6 a month as of 16 October 2026.[20][4]

Which model should I use for an audiobook?

ElevenLabs suggests v4 or Multilingual v2 for audiobooks and narration. Studio still defaults to Multilingual v2, so check Project settings.[2][5]

Can I use a Dubbing v2 track in a streaming series?

Not without written authorisation from ElevenLabs. User-generated platforms and paid ad placements are allowed.[21]

Does ElevenLabs train on my recordings?

Its terms allow it to use your content to improve its services. You can opt out in Data use, but only for future use.[20][27]

The next step

Whichever tool you choose, the review gates matter more than the model. Read our guide to building an AI voiceover workflow with review. For the video side, see our shortlist of AI video generators for creator workflows.

Sources and evidence note

Product features, plans, limits and terms come from ElevenLabs' documentation, pricing pages, changelog, blog, terms and privacy policy. They describe what the company says its products do, not tested performance. Prices are US list prices before tax as displayed on 16 October 2026. Independent evidence comes from the Consumer Reports voice-cloning assessment of March 2025, read in full; its findings describe the products as they were then. The figure and worked example are illustrative. All pages were accessed on 16 October 2026.

Related reading

• Runway review: usable video outputs and editing
• Vapi's approach to voice agents