Correction · October 12, 2026: Website access checks are now explicitly separated from product tests. We corrected the introductory-price description, removed the inconsistent cross-vendor price table and conflicting detection statistics, and narrowed claims about service closures. The original September 19 access observations remain dated historical records.
Contents
- Can You Actually Tell If a Voice Is Fake?
- What Is Voice Cloning AI?
- What the 12 Website Access Checks Show
- What Does Voice Cloning Actually Cost?
- How Does the Technology Actually Work?
- What People Actually Use It For
- The Catch Nobody Mentions: Fraud and Misuse
- FAQ
Voice cloning AI builds a digital copy of one specific person’s voice from recordings of them speaking, then uses that copy to generate new speech in text they never said. It runs on the same family of neural networks behind text-to-speech, just pointed at one person’s pitch, pace and accent instead of a generic voice. The website observations below were recorded on September 19, 2026. They show HTTP and DNS access from one location, not whether a service can produce a good voice clone. We did not sign in, create clones, or compare audio quality.
Can You Actually Tell If a Voice Is Fake?
Recognizing a synthetic voice and proving a caller’s identity are different problems. Three research papers help explain why a convincing demo, a listener’s impression and a detector score cannot be treated as interchangeable evidence.
Claim 1: “A person can hear that a voice is fake”
Mai and colleagues (PLOS ONE, 2023) studied 529 participants using English and Mandarin recordings. In the single-clip task, overall classification accuracy was 70.35%; the share of fake clips correctly identified was 73%. When participants heard a genuine and synthetic version of the same sentence together, identification reached 85.59%. These are different tasks and denominators, not competing estimates of one universal accuracy.
Limit: The study used VITS-generated speech and a restricted set of speakers and languages. It does not measure today’s commercial clones, Korean callers, or a stressed person answering an unexpected phone call. Our assessment: The claim is too broad. Reference audio helped in this experiment, but that does not justify trusting a familiar-sounding caller.
Claim 2: “A strong benchmark score means a detector is dependable”
Müller and colleagues (ASVspoof, 2021) found that leading and trailing silence in the ASVspoof 2019 logical-access data correlated with real-versus-synthetic labels. In their RawNet2 experiment, equal error rate on the 2019 evaluation set rose from 3.61% to 15.50% when silence was trimmed during training. A model can partly learn a dataset shortcut rather than the distinction its score appears to measure.
Limit: This result concerns specific datasets, preprocessing and models; it does not establish that every detector fails. Equal error rate is also not ordinary accuracy. Our assessment: Ask whether evaluation included new speakers, generators and recording conditions. A score on one benchmark is incomplete evidence for a different deployment.
Claim 3: “A detector’s confidence is the probability that this caller is fake”
Pascu and colleagues (Interspeech, 2024) investigated generalization and calibration with self-supervised speech representations and evaluated across eight datasets, including multilingual and partly synthetic audio. Calibration asks whether numerical confidence corresponds to observed outcomes; generalization asks whether performance carries over to unfamiliar data. Improving one reported score is not the same as resolving both questions.
Limit: This is a research evaluation, not certification of a consumer service or a Korean telephone workflow. Our assessment: A detector output needs its threshold, test population and false-positive behavior explained. It should not be presented as a self-sufficient identity verdict.
A worked example: why “99%” can still produce many false alarms
This is an illustrative calculation, not a product measurement or a result from the papers. Suppose 10,000 calls contain 100 fakes (1%). Assume a detector catches 99% of fake calls and wrongly flags 1% of genuine calls. It flags 99 fakes and 99 of the 9,900 genuine calls. Only 99 of 198 alerts are real fakes: 50% of alerts. Changing the underlying rate or either error rate changes that result. The two assumed error rates are not interchangeable with the equal error rate quoted above.
What this changes in practice: A flag should trigger an independent check, rather than an accusation. A reassuring result should not override a request to verify an unexpected payment or account change through a known contact channel. This is our practical interpretation, not a procedure tested by these studies.
What to ask before using a voice detector
- Population: Which languages, speakers, microphones and call codecs were tested? Were they represented in training?
- Errors: At the proposed threshold, how many genuine recordings were flagged and how many fakes were missed? Ask for both counts and denominators.
- Transfer: Was performance checked on unseen generators and altered recording conditions, not only a random split of one dataset?
- Decision: What independent verification follows a flag, and what happens when the detector is uncertain?
Research scope: This is a focused comparison of three papers whose methods or results were consulted on October 12, 2026, not a systematic review or a reproduction of their experiments. The 2021 paper explains an evaluation pitfall; it is not a ranking of 2026 products.
What Is Voice Cloning AI?
It is software that learns the sound of a specific person’s voice from audio samples, then generates new sentences in that same voice — sentences the person never actually spoke.
Resemble AI puts it plainly: the technology “uses artificial intelligence (AI) and deep learning algorithms to replicate a person’s voice with high accuracy.” Soniox frames it a bit differently but lands in the same place, describing it as “the creation of a synthetic voice that reproduces a specific real person, built from samples of their speech.”
What gets captured isn’t just pitch. Timbre, tone, accent, pace and resonance all get modeled. Under the hood, systems compress a voice down into what researchers call a “speaker embedding” — a compact numeric fingerprint of how someone talks, according to Soniox. That embedding is what gets fed into a text-to-speech model later, whenever new audio needs generating.
One thing worth saying up front: this is different from a voice changer or a filter. A clone is trained on a real person. That is also exactly why it raises the fraud and consent questions covered further down.
What the 12 Website Access Checks Show
Before trying a tool, check whether its website is accessible. That is a useful first step, but it cannot establish product functionality, account availability, or audio quality.
Checked 2026-09-19, 01:12 KST, from a single machine in South Korea. Method: curl against each homepage, following redirects, reading the final HTTP status. For the two domains that failed outright, we cross-checked against three independent DNS resolvers to compare DNS responses. This does not establish that a business has closed or that every user is affected.
The recorded check returned an HTTP 200 response for eight homepages and a DNS or HTTP error for four. These are historical observations, not a current availability ranking.
| Service | Status | What we saw |
|---|---|---|
| ElevenLabs | HTTP reachable at check | HTTP 200 |
| Resemble AI | HTTP reachable at check | HTTP 200 after redirect |
| Murf.ai | HTTP reachable at check | HTTP 200, but pricing is JS-rendered |
| Descript | HTTP reachable at check | HTTP 200 after redirect |
| WellSaid Labs | HTTP reachable at check | HTTP 200 after redirect |
| Respeecher | HTTP reachable at check | HTTP 200 after redirect |
| Voice.ai | HTTP reachable at check | HTTP 200 |
| Speechify | HTTP reachable at check | HTTP 200 |
| Play.ht | Access failed at check | No DNS answer on any of 3 resolvers |
| Replica Studios | Access failed at check | SERVFAIL |
| LOVO.ai | HTTP error at check | HTTP 402, Vercel error DEPLOYMENT_DISABLED |
| Coqui | Access failed at check | GitHub Pages 404; this response alone does not establish the reason |
A caveat, and it’s a real one. This was one machine, one moment, one network path. A CDN-backed site can flicker in and out by region. The DNS comparison adds another observation; it does not rule out every local, regional, or temporary cause. Retry the official site before making a purchase decision.
An HTTP 200 response can also be a marketing page or an error screen. To evaluate the actual product, a separate test would need consented voice samples, the same script and language, a record of settings, and listening comparisons. That test has not been performed for this article.
What Does Voice Cloning Actually Cost?
Compare the specific cloning feature and billing period, rather than the cheapest headline price. A free text-to-speech plan is not necessarily a free voice-cloning plan.
Official pricing page checked October 12, 2026: ElevenLabs lists Instant Voice Cloning in Starter and Professional Voice Cloning in Creator. Its monthly Creator price is $22, with $11 displayed as a first-month discount. Taxes are excluded. The earlier version of this article presented $11 without that introductory-price distinction; that was misleading and has been corrected.
The previous cross-vendor price table has been removed because its billing periods, discounts and feature scope were not consistently established. It should not be used as a like-for-like comparison.
- Confirm which clone type is included and whether the speaker must verify their identity.
- Separate first-month discounts from recurring charges, and monthly billing from annual commitments.
- Check included generation allowance, overage charges, commercial-use rights and cancellation terms.
- Try a consented sample in your target language before committing. A website access result is not a substitute for this.
How Does the Technology Actually Work?
Strip away the branding and systems break the problem into two pieces: who is speaking, and what they’re saying. A neural network learns the first part — timbre, pitch, rhythm — from a sample of someone’s voice. That identity then gets fused with new text to generate speech nobody actually said out loud.
Named model families show up across vendor documentation:
- Text-to-speech models: Tacotron 2, FastSpeech, VALL-E
- Vocoders (the part that turns model output into audio): WaveNet, WaveGlow, HiFi-GAN
- Voice conversion models: CycleGAN-VC, StarGAN-VC, VAE-based systems
- Zero-shot speaker-adaptation and embedding models
A typical pipeline, synthesized across vendor write-ups, looks like this:
- Collect audio samples of the target speaker
- Clean and preprocess the audio — noise removal, normalization, trimming silence
- Extract features: pitch, timbre, prosody, spectrograms, even breath patterns
- Train a neural model (RNN, CNN, or Transformer) to map text onto those voice characteristics
- Synthesize new speech from a text input
- Optionally fine-tune for emotion and expressiveness
- Deploy, ideally with safeguards like watermarking or provenance tags
There are two broad approaches to how much audio a clone needs, according to Soniox:
| Approach | Audio needed | Result |
|---|---|---|
| Zero-shot cloning | Seconds | A recognizable timbre, fast |
| Fine-tuning | Minutes to hours | A more faithful, stable result |
Exactly how much audio any single tool requires varies by product and by what “good enough” means for the use case — vendor marketing pages aren’t a reliable source for it.
What People Actually Use It For
Not all of this is scam calls and deepfakes. Potential legitimate uses include accessibility and media production; actual suitability depends on consent, product support and output quality.
- Content creation — YouTube, podcasts, audiobooks
- Dubbing, localization, ADR and alternate takes in film and media
- Voice banking for people with degenerative speech conditions such as ALS
- Restoring a voice that illness took away — Soniox cites Val Kilmer’s voice being restored after throat cancer as an example
- Customer service and IVR — subject to the service’s language support, permissions and actual performance
- Education and tutoring
- Historical voice preservation
- Brand-voice consistency in marketing
- Music production and entertainment
The Catch Nobody Mentions: Fraud and Misuse
This is the part most explainer posts skip entirely, and it’s arguably the more important half of the story.
The FTC ran an exploratory “Voice Cloning Challenge” looking for ways to prevent, monitor and evaluate malicious voice cloning. It distributed $35,000 in prize money across the winning entries, plus one no-prize recognition award. The FTC named three threat categories it’s watching: fraudulent extortion targeting families and small businesses, appropriation of creative professionals’ voices, and misuse of biometric or creative data.
Its own conclusion is blunt. These risks, the agency wrote, “cannot be addressed by technology alone” — enforcement and rulemaking have to do part of the job too.
FAQ
Are Instant and Professional Voice Cloning the same feature? No. ElevenLabs lists them separately in its plan descriptions. Check the required feature, recurring price and speaker-verification requirements on the official site before subscribing.
Does voice cloning work in languages other than English, like Hindi? None of the sources here name Hindi specifically. They do describe voice cloning as generally language-agnostic, since it models a voice’s characteristics rather than any one language’s phonemes — but support varies by tool.
Does a familiar-sounding voice establish someone’s identity? No. Treat an unexpected request for money or account access as a separate identity-verification task. Contact the person through a number or channel you already trust rather than relying only on the caller’s voice.
What remains untested: current access to all 12 sites, account creation, voice similarity, supported languages, latency and generation costs under the same workload. This article does not rank those products or recommend one on the basis of audio quality.
Sources:
- ElevenLabs — What is voice cloning and how does it work with AI?
- Resemble AI — Understanding AI Voice Cloning: What, Why, and How
- Soniox — The Voice AI Wiki: What is voice cloning?
- Federal Trade Commission — FTC Voice Cloning Challenge
- Adaptive Security — The Ultimate Guide to AI Voice Cloning Scams
- Smallest.ai — Understanding What Voice Cloning Is and How It Works
- Historical HTTP/DNS observations recorded 2026-09-19; these observations were not independently repeated for this correction. ElevenLabs official pricing was consulted for the 2026-10-12 price correction.
Image credits
- Photo by Jonathan Velasquez on Unsplash
- Photo by cottonbro studio on Pexels
- ElevenLabs Logo 04 by Ammaarreshi, licensed under BY-SA (via Openverse)