ElevenLabs, Murf, or Cartesia: pick by what the voice has to do
Three AI voice tools that get compared as rivals are built for different jobs: what each bill turns into in finished minutes, whose voice you are allowed to publish, and which one your production actually needs.
Two things landed within two days of each other, and together they change how you should buy an AI voice.
On 28 September, ElevenLabs released v4 and v4 Turbo. Both models read performance direction written into the script, and both clone a voice from ten seconds of audio 12.
On 30 September, the Tokyo District Court held that a person's voice can be protected as a publicity right. It was the first Japanese decision to extend that protection past a name and a likeness 3.
So cloning got cheaper and the right to a voice got clearer in the same week. The voice is now the part of a production that decides what the rest of it costs, and whose name has to be on the permission.
The three tools below get lined up as rivals. Each one is built for a different job, and the job should settle the purchase before any voice samples do.
Quick pick
ElevenLabs — one person or a small studio that wants the largest voice library and the top of the blind-test table. Creative starts at US$6 a month 4.
Murf — a team producing training or course audio on a schedule, where monthly hours and project slots matter more than the last increment of realism. Creator is US$29 a month 5.
Cartesia — a voice inside a product, an app or a phone line, billed by the minute. Pro is US$5 a month 6.
Comparison at a glance
Prices are the vendors' published figures, checked on 27 October 2026. Murf's Creator price is the exception: its own pricing page now opens on its API plans, so that figure comes from Zapier's September test 5.
The last row uses Artificial Analysis's blind tests, where listeners hear two voices without knowing the vendor. The result is an Elo rating: a higher number means the voice won more comparisons 7.
| Field | ElevenLabs | Murf | Cartesia |
|---|---|---|---|
| What you are buying | A production suite for recorded audio: speech, dubbing, audiobooks, music and sound effects 4 | A studio for scripted voiceover that a team shares, with dubbing sold as a second product 8 | Speech models and voice agents behind an API, for audio a product generates live 9 |
| Whose voice you get | 17,500 library voices, an instant clone of your own, or a trained professional clone 2 | A library of professional voice artists, with cloning on the vendor's terms 10 | Library voices, instant clones, and professional clone slots from the US$49 plan upward 6 |
| How you are billed | Credits, one per character of speech, shared with dubbing, music and transcription 4 | Voice generation time in hours, plus project credits that cap how many projects exist at once 8 | Credits for speech at one per character, plus prepaid dollars for agent minutes 6 |
| Entry paid price, as published | Starter at US$6 a month, or US$5 a month on annual billing 4 | Creator at US$29 a month, about US$19 on yearly billing 511 | Pro at US$5 a month 6 |
| What that tier includes | 30,000 credits, a commercial licence, instant cloning, Dubbing Studio 4 | Two hours of voice generation a month and five project credits 8 | 100,000 credits, about 133 minutes of speech, a commercial licence, instant cloning, three concurrent requests 6 |
| What moved most recently | Eleven v4 and v4 Turbo shipped on 28 September 2026, with 90+ languages 1 | Falcon 2, its current model, dates from July 2026 7 | Sonic 3.6, its current model, dates from August 2026 7 |
| The limit that interrupts work | Dubbing spends 2,000 to 10,000 credits a minute; narration runs about 1,000 4 | The free workspace stops at ten minutes and blocks downloads 8 | No professional clone slot below US$49; overage is US$65 per million credits on Pro 6 |
| Independent signal | Joint first on blind tests of each vendor's own voices, Elo 1,333 7 | 26th to 29th on the same board, Elo 1,154 7 | 4th to 5th, Elo 1,292 7 |
ElevenLabs: one studio, two meters
The Creative plans are a single credit pool that every product draws from. Speech costs one credit per character, and ElevenLabs treats 10,000 credits as roughly ten minutes of audio 4.
The tiers run from Starter at US$6 for 30,000 credits to Business at US$990 for six million credits and ten seats. Annual billing charges ten months, which puts Creator at US$18.33 a month in place of US$22 4.
Unused credits roll over for up to two months, so a balance can reach three times the monthly allowance. Cancelling forfeits whatever is left 4.
Dubbing is where the pool drains. Automatic dubbing costs 2,000 credits a minute with the watermark and 3,000 without; the dubbing studio costs 5,000 and 10,000 4.
A ten-minute video therefore takes 20,000 to 100,000 credits. Twenty thousand is two-thirds of a Starter month, and a hundred thousand is a sixth of a Pro month, for one video in one language.
v4 is the reason the pool matters more this quarter. It follows inline direction such as [whispers] or [pause], holds a voice steady across long scripts, and takes 10,000 characters in a single generation 2. Language coverage went from 70 to 90 1.
Cloning comes in two grades. An instant clone needs about a minute of audio plus a consent confirmation. A professional clone needs 30 minutes to three hours of clean recording, plus a verification pass that proves you own the voice 12. Clones trained before v4 need retraining to sit right on the new model 2.
Agents are a second purchase on a second meter. ElevenAgents bills call minutes: 75 minutes on Starter, 1,238 on Pro, 12,375 on Business 13. Extra minutes cost US$0.08, a burst call costs US$0.16, and the language model plus your phone provider bill on top 13.
Running a phone line on the same voice means buying again, on a meter priced per minute. That second purchase is the mistake this pricing structure invites.
On Artificial Analysis's blind tests, v4 and v4 Turbo share first place at 1,333. Alibaba's Qwen-Audio-3.1-TTS-Plus follows at 1,306 and Cartesia's Sonic 3.6 at 1,292 7. The board compares each provider's own voices, so it ranks models.
Zapier spent three weeks on the category and made ElevenLabs its all-in-one pick. Its note was occasional inconsistencies in the output, worst in sound effects 5.
Murf: the studio a team shares
Murf meters voice generation time in hours. Creator carries two hours a month and five project credits, Growth eight hours and 50, Business 20 hours and 200 8.
A project credit caps how many projects can exist at once, so the second number decides whether a 40-lesson course fits. Creator lists at US$29 a month. Yearly plans take a flat 33% off, which puts Creator near US$19 511.
The free workspace lasts ten minutes and blocks downloads, so it is a place to test a voice 8.
Zapier's tester picked Murf for emphasis control: click a word, drag a point up or down, and that word lands harder or softer 5. The same test warns that only some library voices hold up, with many still sounding robotic. It also found the paid voices materially better than the free ones 5.
The blind board shows what that judgment costs. Murf's Falcon 2 scores 1,154, in 26th to 29th place and about 180 points behind Eleven v4 7.
For a compliance module that plays once, that gap costs nothing. For an ad read that has to carry a brand, test the voice before you buy the year.
Murf's studio voices come from named professional voice artists, who are paid royalties when their avatar is used. The company also says it will back commercial customers against copyright claims over those voices 10. That arrangement puts the consent question on the vendor, which counts when the recording goes into a client deliverable.
Dubbing is a separate product. Murf Dub gives 100 free minutes on sign-up, up to 40 target languages, a native-speaker review step and MP4 export 14.
The API is priced by character: US$0.01 per 1,000 characters on Falcon 2.1 and US$0.03 on Gen2. Translation runs at US$0.02 per 1,000 characters and speech to text at US$0.24 an hour 15. Murf states SOC 2 Type II, ISO 27001, HIPAA and GDPR compliance on the same page 15.
Cartesia: the voice inside a live product
Cartesia ships as infrastructure. Audio comes back through an API into your app, your call centre or a device 9.
Speech is metered a credit a character. The plan page converts that into minutes. 20,000 credits on the free plan is about 27 minutes, 100,000 on Pro about 133, 1.25 million on Startup about 1,667, and eight million on Scale about 10,667 6.
Agents carry their own budget. Each plan includes prepaid dollars equal to its monthly price for agent minutes at US$0.06 a minute, and a Cartesia-provided phone number adds US$0.014 a minute. That buys 17 minutes on the free plan and roughly 83 hours on Scale 6.
Professional clone slots are the real tier gate: none on Free or Pro, two on Startup at US$49, four on Scale. Training a clone is free, and speech generated with it costs the same credit per character as any other voice 6.
Cartesia charges overage at US$65 per million credits on Pro, US$45 on Startup and US$38 on Scale. Concurrency starts at two requests on Free and rises to three on Pro, five on Startup and 15 on Scale 6.
Cartesia's pitch is latency, with its models running in-region, on-premise or on-device 9. ElevenLabs answers with v4 Turbo, which it publishes at about 100 milliseconds of inference and 150 milliseconds to first speech 2.
Sonic 3.6 sits at 1,292 on the blind board, fourth or fifth, 40 points behind Eleven v4 and ahead of Google's Gemini 3.8 Flash TTS at 1,280 7. At that distance the deciding factors become the per-minute rate and where the audio has to run.
If all you need is other languages
Before buying dubbing minutes, check the free switch you may already have. YouTube generates automatic dubs for eligible channels by default and marks the video as auto-dubbed in its description 16.
From an English original, YouTube dubs into 20 languages, among them Spanish, Portuguese, Hindi, Indonesian, Japanese and Korean. From most other source languages it produces English; Spanish also gets Arabic and Portuguese, and Korean also gets Indonesian 16.
The boundaries are specific. Videos over 120 minutes, videos with little speech or music only, videos whose source language YouTube cannot identify, fast speech, and anything under a copyright claim are ineligible. Automatic dubs cannot be edited 16.
Two settings change the deal. Manual review lets you approve each dub before it publishes, and the lip-sync feature is in early access for selected channels 16.
Paid dubbing earns its place in three situations. The audio has to live off YouTube, in an ad, a course platform or a client's site. One brand voice has to stay identical across languages. Or the script is dense with medical, legal or product names that a machine translation will mangle.
Then the paid meters come back, and the arithmetic from the ElevenLabs section decides the bill. A ten-minute video dubbed by machine costs 20,000 credits at the low end and 100,000 at the high end 4.
Whose voice you may use
Every recording you upload falls into one of four categories. It is your own voice, a voice you hired and have permission to reproduce, a library voice the vendor licenses to you, or a designed voice that imitates nobody.
ElevenLabs draws the hardest line. A professional clone requires you to verify that you own the voice being cloned. The check is a fresh recording compared against the long sample you supplied, and an instant clone asks you to confirm consent instead 12. The company also says its speech classifier can flag generated audio as AI-made 2.
Murf works the other way round, selling voices it has already cleared. Its artists are paid royalties on use, and the company offers commercial customers support against copyright claims over its voices 10.
Cartesia lists instant and professional cloning as features of its paid plans. Its pricing page states the clone slots and leaves the consent process to the account holder 6.
The law moved on 30 September. The Tokyo District Court held that a voice, like a likeness, is a symbol of a person's personality. Unauthorised use infringes publicity rights where the use aims solely at exploiting the voice's drawing power 3.
The court dismissed the claim itself on procedural grounds, because the videos had already been taken down. It set no damages and no test for how similar a clone has to be 3.
Japan's Ministry of Justice reached a compatible view in August 2026, in a study-group report. Voices that identify a specific person are legally protected, and people hold a right to be asked before their voice is used 3.
For a buyer, the practical rule is paperwork. Get permission in writing that names the uses, the languages, the term and the platforms. A voice actor's consent to one campaign leaves the clone that outlives it uncovered 3.
Pick by the job
- You narrate your own videos and scripts. ElevenLabs Starter, US$6 for 30,000 credits, which is about half an hour of finished speech a month 4.
- Several people produce training or course audio every month. Murf's Business workspace, for 20 hours of generation and 200 project slots. Creator, at US$29 a month, covers two hours and five projects 58.
- One video has to exist in five languages. YouTube's free automatic dubs first, if it publishes there 16. Off YouTube, ElevenLabs automatic dubbing at 2,000 credits a minute, or Murf Dub's 100 free minutes 414.
- A customer has to hear a voice answer now. Cartesia Startup at US$49, which brings professional clones and five concurrent requests 6. Choose ElevenAgents at US$0.08 a minute when the voice has to match the one on your videos 13.
- An audiobook in your own voice. ElevenLabs' professional clone on Creator or above, with 30 minutes to three hours of clean recording to prepare first 12.
Traps that waste the money
- Reading the annual rate as the monthly one. ElevenLabs' yearly billing is ten months' worth, so Creator shows US$18.33 against US$22. Murf takes 33% off yearly 411.
- Dubbing on a narration budget. At 2,000 to 10,000 credits a minute, one dubbed video costs two to ten times what narrating the same script does 4.
- Buying a creative plan for a phone agent. ElevenAgents runs on call minutes, and the language model bills separately from the voice 13.
- Judging quality on a free tier. Murf's free workspace stops at ten minutes and blocks downloads. Zapier found the paid voices materially better than the free ones 58.
- Assuming a clone carries over. Clones trained before v4 need retraining before they sound like themselves again 2.
- Expecting professional cloning on the cheap plans. At Cartesia, the professional clone slot starts on the US$49 plan 6.
- Letting overage set your tier. Cartesia charges US$65 per million credits on Pro and US$38 on Scale. Past the allowance, the cheaper plan costs more 6.
Test it in one afternoon
- Take one real 200-word script you have already published.
- Run it through two candidates on their cheapest paid tier, with the same voice settings.
- Count the words you had to regenerate because of pronunciation or emphasis. That count is your quality score.
- Note what the test consumed, then price a month at your real volume: minutes of speech for narration, minutes of calls for an agent.
- Listen on a phone speaker, where most of your audience will meet the voice.
- Confirm the licence covers the work you sell, whether that is client video, paid ads or a paid course.
- Only then subscribe, on monthly billing, and switch to yearly once a second project has come through.
Bottom line
ElevenLabs if the audio has to be the best-sounding thing you ship and you can live inside a credit pool. Murf if a team shares the work and the volume is counted in hours. Cartesia if the voice has to answer while the customer is still on the line.
Prices were checked on 27 October 2026, and the blind-test scores read from Artificial Analysis's provider-voice board on the same date.
References
- 1
- 2Eleven v4 and Eleven v4 Turbo
elevenlabs.io
- 3
- 4ElevenLabs pricing
elevenlabs.io
- 5The 8 best AI voice generators in 2026
zapier.com
- 6Cartesia pricing
cartesia.ai
- 7Provider Voice Arena leaderboard
artificialanalysis.ai
- 8Murf: what is a workspace
help.murf.ai
- 9Cartesia
cartesia.ai
- 10Murf ethical AI
murf.ai
- 11Murf: additional voice generation time
help.murf.ai
- 12Instant vs. professional voice cloning
elevenlabs.io
- 13ElevenAgents pricing
elevenlabs.io
- 14Murf AI dubbing
murf.ai
- 15Murf pricing
murf.ai
- 16Use automatic dubbing
support.google.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.