Voice cloning can look like a simple monthly subscription, but the plan price is only one part of the bill. You also need to count generated characters or minutes, retries, commercial-use needs, seats, and the time spent reviewing the result. There is another change worth knowing before comparing vendors: Resemble AI no longer sells voice generation as a commercial product; it now focuses on detection, watermarking, and media security.
What ElevenLabs currently charges for voice cloning
ElevenLabs’ public monthly plans bundle voice-cloning access with a shared pool of credits for text-to-speech and other products. The entry plan includes Instant Voice Cloning; Professional Voice Cloning starts at Creator. These are subscription tiers, not a separate one-time fee for creating one voice.
| Plan | Listed monthly price | Voice-cloning access shown | Credits and seats |
|---|---|---|---|
| Starter | $6 | Instant Voice Cloning | 30,000 shared credits |
| Creator | $22 | Professional Voice Cloning | 121,000 shared credits |
| Pro | $99 | Includes Creator features | 600,000 shared credits |
| Scale | $299 | 3 Professional Voice Clones | 1.8 million credits; 3 seats |
| Business | $990 | 10 Professional Voice Clones | 6 million credits; 10 seats |
The price page currently also shows introductory offers, including a temporary Starter promotion and a discounted first month for Creator. Its plan FAQ lists the regular monthly prices above; check the checkout screen for the exact offer, renewal price, tax, and billing interval before subscribing. Prices and promotions can change.
Estimate the bill from the work, not just the plan name
ElevenLabs says that credits are shared across products. Text-to-speech generally uses credits per text character, with the exact rate depending on the model: its pricing FAQ lists one credit per character for V2 Multilingual, and 0.5–1 credit per character for certain Flash/Turbo models through the API. Other products draw from the same pool, so speech-to-text, dubbing, voice changing, or sound effects can leave less for speech generation.
A useful first-pass calculation is:
Monthly plan cost ÷ expected usable speech output
Estimate characters for the chosen model, then add expected rewrites, regenerations, other products using the shared credits, and the cost of any seats you need.
For example, 10,000 characters generated with a model that charges one credit per character would use about 10,000 credits before other products or retries. That fits within the listed 30,000-credit Starter pool, but it does not mean every 10,000-character project costs the same: model choice, language, output length, re-generation, audio quality, and bundled products can change usage. Treat the provider’s calculator and your own sample run as the budget check.
Do not compare a “minute” from one service with a “credit” from another unless you have tested the same script, language, output settings, and quality. A minute of generated audio can contain very different amounts of spoken text, and a subscription may include tools unrelated to cloning.
Why the old Resemble AI price comparison is no longer valid
Resemble’s current product notice says it has stopped offering voice generation commercially and is focused on AI detection, watermarking, and multimodal security. Its current pricing page sells detection and security services: for example, Flex is pay-as-you-go, while Team and Business are subscriptions for teams. Those prices are not voice-cloning plan prices and should not be compared with ElevenLabs’ generation plans.
Resemble says its voice models remain open and free for research and development. That is a different use case from buying a hosted, supported commercial voice-cloning service. If you need a managed service, confirm that the vendor currently offers generation in your region and plan before relying on an old comparison table.
Can people or detectors reliably identify a cloned voice?
A 2025 Scientific Reports study found that participants matched an AI voice clone’s identity to the real speaker about 80% of the time and correctly identified AI-generated voices only about 60% of the time in the study’s tasks. That is a warning against relying on “it sounds like them” or “I can hear the fake.” It is not a universal detection rate for every language, model, clip, or real-world call.
Automatic detection has a different limitation: a detector can perform well on generators and recording conditions it has seen, then struggle when the generator, codec, noise, or speaker changes. A 2025 AAAI paper studies this cross-domain generalization problem. Its results support testing detectors on unfamiliar generators and realistic audio, rather than treating one vendor’s accuracy claim as a guarantee.
Watermarking is another approach, but it answers a narrower question. Some research embeds a mark in source speech to make later speaker adaptation harder; other systems mark generated audio so a detector can look for provenance. A mark can help when it is present and survives processing, but an absent or undetected mark does not prove that a recording is human or authentic.
| Method | What it can contribute | What it cannot establish alone |
|---|---|---|
| Human listening | May notice unusual phrasing, timing, or context | Reliable identity verification; the 2025 study shows people can mistake clones for the source voice |
| AI-audio detector | A risk signal for a specific tool and tested conditions | Universal proof; performance may shift on new generators, codecs, languages, and noise |
| Watermark or provenance record | Evidence that a compatible system marked or registered media | Proof that an unmarked recording is genuine, or proof that the content itself is true |
Use a voice clone responsibly—and protect against impersonation
- Use a voice you control: ElevenLabs’ Professional Voice Cloning requires verification that the voice belongs to the account holder; its documentation says you cannot create a Professional Voice Clone of another person, even with their consent. Other products and services have their own consent checks and terms, so read the exact workflow.
- Keep permission specific: Before using anyone’s voice, document what may be cloned, where it may be published, how long the source recordings and model may be retained, who can access them, and how to revoke access. A vendor check is not a substitute for your own rights and privacy review.
- Tell listeners when it matters: Label synthetic speech in contexts where a listener could reasonably mistake it for a real person, especially endorsements, news, customer contact, or sensitive instructions. Check the rules that apply in your jurisdiction and distribution channel.
- Verify urgent requests out of band: If a call asks for money, credentials, or a sensitive action, hang up and call the person using a number already on file. A second channel or agreed passphrase is more useful than debating whether the voice sounds real.
- Do not make a detector the final decision-maker: Treat its score as one signal. Confirm identity through an independent channel and preserve the original file and context for review.
The practical takeaway
For a creator choosing a hosted cloning service, ElevenLabs currently bundles Instant Voice Cloning in Starter and Professional Voice Cloning from Creator; the subscription also buys a shared credit pool, not unlimited output. Resemble AI has pivoted away from commercial voice generation, so its current detection subscription is not a like-for-like alternative.
For budgeting, estimate output in the actual language and model you plan to use, include retries and shared-credit use, and verify the renewal price. For trust and safety, assume a convincing voice may be synthetic: consent and clear labeling help on the creation side, while independent verification and layered detection help on the receiving side.
Sources and research
- ElevenLabs current pricing, plan credits, and credit-use FAQ
- ElevenLabs Professional Voice Cloning: training samples and voice verification
- ElevenLabs safety controls and AI Speech Classifier
- Resemble AI product notice: commercial voice generation discontinued
- Resemble AI current detection and security pricing
- Barrington, Cooper & Farid (2025), “People are poorly equipped to detect AI-powered voice clones,” Scientific Reports
- Ren et al. (2025), “Improving Generalization for AI-Synthesized Voice Detection,” AAAI
- Zong et al. (2025), “AudioMarkNet: Audio Watermarking for Deepfake Speech Detection,” USENIX Security
- Resemble AI ethics and detection limitations
Checked: October 13, 2026. Prices, promotions, plan features, and product availability can change. Research findings apply to their study settings and do not establish universal detector or listener accuracy. We did not generate cloned audio or independently benchmark a detector for this article.