ElevenLabs Review (2026): Eleven v4, Audio Tags, Pricing & Voice Cloning

Last updated November 10, 2026

ElevenLabs’ Eleven v4 gives creators more control over how a line sounds, from the pause before a punchline to the shift from a quiet aside to a louder reaction. It also brings Professional Voice Clones back to its expressive model and raises the limit to 10,000 characters per generation. Eleven v4 Turbo brings audio-tag control to lower-latency voice workflows.

This review covers when I’d reach for v4, what happened in a short two-voice comedy test, and how to pick a plan around the audio you actually finish.

Affiliate disclosure If you sign up through a link on this page, I may earn a commission at no extra cost to you.

Jump to What Matters

• Quick answers first
• What’s new in Eleven v4 and v4 Turbo
• Audio tags and a dialogue test
• Dubbing and localization
• Is ElevenLabs worth it?
• Pricing and credits
• Commercial use and legal questions
• Voice cloning questions
• Why your ElevenLabs output sounds weird
• API and technical questions
• Best for / not best for
• Biggest mistakes and fixes
• Here’s the exact way I’d use ElevenLabs
• FAQ

Quick Answers First

If you only want the short version, here it is.

ElevenLabs is one of the strongest AI voice tools for realistic text-to-speech, voice cloning, API-based voice workflows, and long-form audio production. It has a free plan, but commercial rights are tied to paid plans, not the free tier. Credits are deducted when you generate audio, not when you download it. Unused credits can roll over, but only up to a limit and only while you stay on the same subscription. Professional Voice Clone is only for your own voice, and if your output sounds robotic or unstable, the cause is usually weak source audio, a poor language-accent match, or too much testing on the wrong voice.

If that already sounds like your workflow, try ElevenLabs here.

What’s New in Eleven v4 and v4 Turbo?

Eleven v3 already supported audio tags and multi-speaker dialogue. The reason to try v4 is how it handles those jobs. ElevenLabs reports more reliable tag following, a wider emotional range, improved speaker consistency across regenerations, and support for Professional Voice Clones. The vendor lists 90+ languages and up to 10,000 characters per generation for v4, compared with 70+ languages and 5,000 characters for v3. Eleven v4 model page

For a short video, I’d start with v4. For an audiobook already in production, I’d render the same short passage in v4 and the existing model and compare voice continuity, pacing, and pronunciation before switching mid-book. For a live agent, I’d time the entire exchange with v4 Turbo, from when the speaker stops talking to when the response is audible. ElevenLabs model documentation

ElevenLabs also says Instant and Professional Voice Clones made before v4 may need retraining for the new model. Its v4 page describes opening the voice and using the plus control to retrain. Keep an approved sample from the previous model so you can compare the same words in both versions.

How Eleven v4 Audio Tags Work

Write a short direction in square brackets before the line it should affect. V4 can interpret natural-language cues for emotion, delivery, pace, and reactions. The same tag can land differently depending on the voice and the surrounding words, so I’d start with one cue at a time and listen for what it actually changed.

• Emotion Try [curious], [mischievously], or [excited].
• Delivery Try [whispers] or [shouts] when the change in volume belongs in the performance.
• Timing Try [pause] or [long pause], then trim or place the silence on your timeline if the cut needs exact timing.
• Reactions Try [laughs] or [sighs] where a reaction belongs. For comedy, I’d leave room for the actor’s reaction before deciding it needs a laugh.

V4 uses audio tags instead of SSML break tags. For unusual names or technical terms, check the pronunciation dictionary support for the workflow you’re using. The vendor also documents combining cues inside a single bracket, which is useful once you know how the selected voice responds to simpler directions. ElevenLabs audio tag examples

What Happened in My Two-Voice Comedy Test

I tried v4 on a short Maggie and Dave refrigerator exchange. The voices sounded more like people talking to each other, and the test used 283 of the 10,000 free credits. I still wanted tighter timing and more emphasis on “vegetables” and “listen.” In a follow-up attempt, the output only gave me the first line for each speaker. That’s one test result. It doesn’t establish a two-line cap.

I’d use the combined dialogue to audition the chemistry. When the next step is lip-syncing two people separately, I’d generate each speaker’s lines as separate clips using the same voice and model, then assemble them on the timeline. If a line needs a half-beat before the comeback, I’d place that gap in the editor so a retake doesn’t change the whole scene.

A Short Test for Timing and Emphasis

1. Render the two lines without tags as a baseline.
2. Try a single [matter-of-fact] or [dryly] cue at the start of a line. Keep character names in the speaker fields where the interface supports them.
3. Keep the version with the best delivery. If the pause still feels too long or too short, adjust it in the editor. Regenerate only the line whose delivery needs work.
4. Track the credits used for the approved scene, including charged retakes. That’s the number to use when estimating a month of episodes.

Developers can use the Create Speech or Stream Speech endpoints with eleven_v4. ElevenLabs also documents Create Dialogue and Stream Dialogue for multi-speaker scripts. Use the endpoint’s structured speaker inputs and voice IDs. Use eleven_v4_turbo where supported when low-latency interactive speech is the job. Speech and dialogue documentation

Is ElevenLabs Worth It?

Yes, for the right person.

If you make YouTube videos, courses, ads, audiobooks, podcasts, app audio, or client voiceovers, ElevenLabs has a real case. The tool is built around text-to-speech, voice creation, cloning, dubbing, Studio-style long-form workflows, and API use. That makes it much more than just a “fun AI voice generator.”

Where people get frustrated is usually not quality first. It is the workflow and pricing model.

The real question is not “does ElevenLabs sound good?”
The real question is “will this save me time without making me burn through credits every time I test a line?”

That is the part a lot of people misunderstand before subscribing.

That is the real tradeoff.

If you already know you want to test it, this is the ElevenLabs link I recommend.

Pricing and Credits

Is ElevenLabs free?

Yes. ElevenLabs has a Free plan, and on the public pricing page it sits alongside Starter, Creator, Pro, Scale, Business, and Enterprise options. The free tier is useful for testing, but it is not the same thing as a real business workflow because the commercial-use rules are different and the included usage is limited.

What do the paid plans unlock?

The public monthly pricing page showed Free with 10,000 credits, Starter at $6 with 30,000 credits, Creator at $22 with 121,000 credits, and Pro at $99 with 600,000 credits when I checked. Creator can show a lower first-month price. Use the renewal price for your budget. Starter includes a commercial license and Instant Voice Cloning. Creator adds Professional Voice Cloning. The v4 launch page lists access to the standard v4 model even on Free, subject to the plan’s usage and licensing rules. Current ElevenLabs pricing

ElevenLabs shares the credit pool across tools. At 283 credits for a generation like my test, 30,000 credits would cover about 106 equally sized charged attempts before any other use. A real episode may require multiple takes, so count the credits per approved scene before choosing a plan.

Does ElevenLabs charge by word, character, or minute?

Text-to-speech credits generally depend on input length and the model. The pricing page documents different rates for some Flash and Turbo API models. Other products can charge by time or generation type. Check the quoted usage for your selected tool, especially if you also generate dubbing, music, or sound effects from the same credit pool.

Do ElevenLabs credits roll over?

Yes, but not forever. The docs say credits reset monthly and unused credits roll over for up to two months. That is useful, but it is not a permanent savings bank where you can quietly stack credits forever and come back later.

Do you lose credits every time you hit Generate?

The initial generation consumes credits. Eligible dashboard regenerations can be free when the text and parameters stay the same and the interface shows an available regeneration. Changing the line, voice, or settings can start a new charged generation. Budget API retries as charged unless the relevant endpoint explicitly says otherwise.

Is there a true preview mode without using quota?

A new read of your script consumes quota. You can audition existing voice samples before generating. For a custom performance, test one or two sentences before rendering a chapter or scene.

How much text can ElevenLabs handle at once?

The website’s Text to Speech tool supports up to 10,000 characters per generation with Eleven v4 and up to 5,000 with the other website models, according to the current help page. API limits vary by model and endpoint. For books or courses, use Studio or work in chunks you can review and regenerate independently. Current generation limits

Commercial Use and Legal Questions

Can I use ElevenLabs commercially?

Yes, on paid plans. ElevenLabs states that the free plan does not include a commercial license and cannot be used for commercial purposes. It also says all paid plans include a commercial license, provided you are not using Beta Services and your use follows its terms, policies, and applicable law.

This is one of the biggest buying questions, and honestly, it should be.

“Sounds good” is not enough.
“Can I safely use it for monetized content or client work?” is the real question.

What if I use the free plan?

ElevenLabs says content created on the free plan or while signed out must be attributed and cannot be used commercially. So the free tier is fine for exploring the platform, but not something I would build a serious client or monetized creator workflow around.

What happens to my content if I cancel?

This is one of the better parts of the policy. ElevenLabs says that once your paid subscription ends, you still keep the commercial license for whatever you generated during that paid subscription. It also says that audio generated outside of a subscription still requires attribution.

So no, you do not lose the right to use properly licensed content just because you stop paying later.

Will my old content always stay in my account?

Not guaranteed forever. ElevenLabs says the content remains accessible for an unspecified amount of time, but it does not guarantee that it will always stay available under the free tier after your subscription ends. So if something matters, save your working files and exports.

Voice Cloning Questions

Can I clone my own voice?

Yes. That is one of ElevenLabs’ main use cases, and the product is structured around My Voices, Instant Voice Cloning, and Professional Voice Cloning.

What is the difference between Instant Voice Cloning and Professional Voice Cloning?

In plain English: Instant Voice Cloning is the quicker, lighter option. Professional Voice Cloning is the more serious version with verification, stricter rules, and better long-term use for people who want a higher-quality branded voice setup.

Can I create a Professional Voice Clone of someone else’s voice?

No. ElevenLabs states that you can only create a Professional Voice Clone of your own voice. Even with consent, you cannot create a Professional Voice Clone of someone else’s voice. The platform requires a verification process to confirm that the voice belongs to you.

That is an important reality check because a lot of people assume “with permission” automatically means “allowed.” For PVC, it does not.

Can someone else share their voice with me?

Yes, but the proper way is for that person to create and verify their own Professional Voice Clone, then share it with you privately using ElevenLabs’ sharing tools.

How much audio do I need for cloning?

ElevenLabs’ voice-cloning guidance says Instant Voice Cloning can work with short audio, while Professional Voice Cloning is designed around a much more serious training set. In the help content, ElevenLabs recommends higher-quality source files and much longer, cleaner audio for Professional Voice Cloning than for Instant Voice Cloning.

Minimum does not equal optimal.

If your sample is noisy, inconsistent, compressed badly, or emotionally all over the place, the clone can come out thin, robotic, unstable, or just plain odd.

How many Professional Voice Clones can I have?

ElevenLabs says PVC slots vary by plan. Free and Starter have no PVC slots. Creator, Pro, and Scale include 1 PVC slot. Business includes 3. Enterprise is custom. If you downgrade below Creator, the PVC may stay in your library, but you will not be able to use it until you upgrade again.

Who has access to my cloned voices?

By default, your voices are tied to your account unless you explicitly share them or publish eligible ones. ElevenLabs also says that only human-verified Professional cloned voices can be shared in the Voice Library. Instant Voice Clones and synthetic voices created with Voice Design are not shareable there.

What is the Voice Library?

The Voice Library is the place where eligible voices can be shared and discovered. For creators, it matters because it is both a shortcut to strong voices and a clue about what kinds of voices are being positioned for wider use.

Can I export my cloned voice?

No. ElevenLabs’ help content says voice clones are usable on ElevenLabs and cannot be exported as a clone asset for use somewhere else. You can save generated audio, but not export the clone itself as a portable voice model.

If voice cloning is the main reason you are considering the tool, check out ElevenLabs here.

Why Your ElevenLabs Output Sounds Weird

This is the section most people actually need.

Why does my clone sound robotic?

Usually because the source audio was the real problem.

ElevenLabs’ troubleshooting guidance points to quality and stability issues when a voice starts whispering, changing tone, breaking, or behaving unpredictably. It also notes that these problems depend heavily on the voice used, how well cloned it is, the quality of the original samples, and how wide the dynamic range is.

If your output looks weird, it’s usually because:

• the source audio was noisy
• the speaking style was inconsistent
• the voice was trained in the wrong language or accent context
• you started testing long scripts before validating the clone on short lines

Fix: re-record cleaner source audio, keep delivery consistent, match the voice to the target language, and test short lines first.

Why does the voice change accent or language?

ElevenLabs’ help content specifically calls out language switching and accent-related issues as real things users can run into, especially in longer generations. It also recommends using the right voice and long-form tools like Studio when quality starts to wobble.

In plain English: just because the tool can push a voice across languages does not mean it will sound natural doing it.

Why does the voice start whispering, changing tone, or breaking?

ElevenLabs says that when a voice drops in volume, starts whispering, or starts distorting, it is most likely a stability issue or a voice issue. It also recommends using Studio for anything longer than a few hundred characters if you are running into these problems, because Studio lets you regenerate specific sections instead of the whole thing.

Why are names, dates, symbols, acronyms, or numbers pronounced badly?

ElevenLabs’ text-to-speech docs specifically call this out as a common issue. Those inputs can be hard because there are often multiple valid ways for the AI to interpret them.

Fix: rewrite tricky phrases phonetically when needed, split complex lines into smaller chunks, and use pause controls or phoneme tools where appropriate instead of hoping punctuation alone will fix everything.

Can I control pauses better?

In v4 use [pause], [long pause], or natural punctuation. V4 does not interpret SSML break tags. Some older models support SSML breaks, so check the model before reusing a script written for another engine. For exact lip-sync or a hard edit, place the silence on the timeline.

Should I use Studio for long scripts?

Yes, usually. ElevenLabs repeatedly points people toward Studio for longer-form work and for troubleshooting issues that get worse in bigger blocks of text. That is one of the simplest upgrades you can make if you are getting decent short generations but messy long ones.

API and Technical Questions

How do I get started with the ElevenLabs API?

The official path is straightforward: create an account, generate an API key in the developer area, and use that key in your requests. If you are eventually planning to connect ElevenLabs to an app, workflow tool, automation, or custom product, that is where you start.

What are the most common API mistakes?

The usual ones are not glamorous. Wrong API key. Wrong voice ID. Wrong model. Or using the default model in one place and assuming it matches another place. That is why “the API sounds different than the website” is such a common complaint.

Why does ElevenLabs sound different in the API than on the website?

One official reason is model mismatch. ElevenLabs’ API help says that if you do not explicitly set a model, your API call may not be using the same model as the website flow you are comparing it to.

This is not a bug most of the time.

It is usually a workflow mismatch.

How do I find the voice ID?

ElevenLabs says the easiest method is to open My Voices, click the three-dot menu for the voice, and copy the voice ID. It also says you can get a list of voices via the List Voices API endpoint.

How do I find the model ID?

ElevenLabs says model IDs can be found via the /v1/models endpoint. Its help content also lists model IDs and notes cost differences between faster/cheaper models and other higher-cost options.

What models should I care about?

Start with the model comparison above. For API work, record the model ID, voice ID, settings, and output format for every approved project so you can reproduce the setup later.

How many languages does ElevenLabs support?

ElevenLabs lists 90+ languages for v4 and v4 Turbo. The count differs by model and product. Test the target language with the specific voice you plan to ship, including names, accents, and any code-switching.

Dubbing, Localization, and Review

Dubbing is a separate workflow from generating an expressive read in v4. ElevenLabs documents a project-based Dubbing v2 API that can translate audio or video into 90+ languages, keep transcripts and translations editable, and regenerate only the regions you change. I’d verify proper nouns, meaning, speaker assignments, and timing with a fluent reviewer before publishing the localized version. Dubbing v2 API release notes

What Other Users Bring Up

G2 reviewers often highlight realistic voices and ease of use while pointing to credit pressure on long scripts and repeated edits. Trustpilot includes praise for the voice library and individual reports of billing or support trouble. Creator discussions also show how quickly retry costs can become the issue. I’d use the current pricing page for quotas, since older reviews can quote different plan figures. G2 reviews · Trustpilot reviews

Best For / Not Best For

ElevenLabs is best for:

• YouTube narration
• branded voice cloning
• multilingual audio
• API workflows
• long-form audio production
• client work that needs realistic voice quality
• creators who want a serious voice platform, not just a novelty generator

ElevenLabs is not the best fit for:

• people who want endless messy retakes without watching credits
• people who expect a free-plan workflow to behave like a business plan
• people who only need a few quick throwaway voiceovers and do not care much about premium quality
• people who refuse to test source audio and then blame every bad output on the tool

Biggest Mistakes and the Fix

Mistake 1: using the free plan like a business plan

Fix: move to a paid plan before using the output in monetized content or client work.

Mistake 2: generating full scripts before testing the voice

Fix: test one or two lines first, because the charge happens when you generate, not when you download.

Mistake 3: blaming sliders when the source audio is bad

Fix: improve the raw sample first. That is usually the real bottleneck when a clone sounds unstable, whispery, or broken.

Mistake 4: expecting one voice to sound native in every language

Fix: use the right voice and workflow for the target language, especially for long-form content.

Mistake 5: assuming the API should magically match the website

Fix: explicitly check the model, voice ID, and settings before comparing results.

Here’s the Exact Way I’d Use ElevenLabs

I would not start with a long script.

I would start with two short test lines:

• one normal sentence
• one tricky sentence with a name, number, or acronym

Then I would lock in:

• the voice
• the model
• the language or accent fit
• the pacing style

Only after that would I move into the full narration.

Why? Because a lot of “bad tool” complaints are really “bad test workflow” problems.

For longer projects, I would move into Studio instead of forcing everything through a single basic generation box. That is exactly the kind of workflow shift that tends to make the tool feel better instead of more expensive and frustrating.

That is the exact reason I would tell a creator to start with ElevenLabs here, test short, then scale up.