How to Detect an AI Voice in 2026: Free Detectors, Warning Signs and What Actually Works

How to Detect an AI Voice in 2026: Free Detectors, Warning Signs and What Actually Works

Short answer: you can't reliably hear the difference anymore. The best 2026 voice models breathe, pause and stumble like people. To check a recording, upload the original file to two different AI voice detectors and look for a watermark or Content Credentials. To check a live call that asks for money, don't analyze the voice at all: hang up and call the person back on a number you already know.

We first published this guide in June 2024. A lot has changed since then: one of the three detectors we recommended no longer exists, another now admits it can't reliably catch its own newest model, and the EU started requiring deepfake labels in August 2026. Everything below was checked on the tools' own pages on 15 November 2026. We did not re-run our own test clips for this update, so we report each tool's documented limits rather than accuracy percentages.

Disclosure: This post may contain affiliate links, and if you decide to buy any of the promoted products, I may receive a commission at no additional cost to you. By doing this, I might feel more inspired to continue writing on this blog. You can read our affiliate disclosure in our privacy policy.

What changed since 2024

  • The PlayHT voice classifier is gone. PlayHT is no longer available, so that tool can't be used.
  • Generator-specific detectors fall behind their own models. The ElevenLabs AI Speech Classifier now states that it "does not reliably classify" audio made with the Eleven v3 model.
  • Watermarks are now checkable by anyone. Gemini can tell you whether audio carries Google's SynthID watermark, and many tools read Content Credentials (C2PA), a kind of signed history attached to a file.
  • Labels are becoming law. In the EU, the transparency rules in Article 50 of the AI Act, including labelling of deepfake audio, apply from 2 August 2026.
  • Voice scams went mainstream. The FBI has warned that criminals use AI-generated voice messages to impersonate senior US officials and personal contacts.

Is it a live call? Verify the person, not the voice

Detectors work on recordings. During a phone call there is no time to upload anything, and real-time voice cloning is good enough that "it sounded exactly like her" proves nothing. Use these checks instead:

  1. Hang up and call back on the number saved in your phone, not the one that called you. Scammers can spoof caller ID.
  2. Ask for your family code word, or a question only the real person could answer and that isn't on social media.
  3. Slow down. Urgency, secrecy ("don't tell Mom") and unusual payment methods (gift cards, crypto, wire transfers, a courier picking up cash) are the scam, whatever the voice sounds like.
  4. Move to a different channel. If someone who claims to be your boss or a government official asks you to switch to another messaging app, that is a pattern the FBI describes in its warning about AI voice messages.

In the US, report attempts to the FBI at ic3.gov. Since February 2024, the FCC has treated AI-generated voices in robocalls as "artificial" under the Telephone Consumer Protection Act, so AI-voice robocalls without consent are illegal.

Free AI voice detectors in 2026

Use these for recordings: voicemails, voice notes, clips from social media, interview audio. Limits are as shown on each vendor's site on 15 November 2026.

ToolWhat it detectsFree optionPaid plans
AI Voice DetectorAI speech from any generator, with a score per segment50 checks a month, clips up to 2 minutes, no cardStarter $49/month (1,000 checks, API); Business $199; Enterprise $599. Yearly billing is cheaper.
ElevenLabs AI Speech ClassifierOnly audio made with ElevenLabsFree, no login; analyzes the first minuteNone
Gemini verification (SynthID)Google's SynthID watermark and Content CredentialsFree with a Google account; audio under 1 hour, up to 100 MBNone
Hiya Deepfake Voice DetectorAI voices in audio and video playing in ChromeFree Chrome extensionNone for individuals
Resemble AI DetectSynthetic audio, video and imagesFree Chrome extension (beta) with a free account, 4 scans a dayBusiness API; see Resemble's pricing page
Deepfake DetectorAudio, video and images50 checks a month, clips up to 2 minutesSame plan structure and prices as AI Voice Detector

AI Voice Detector

The closest thing to a general-purpose free checker for ordinary people. Upload an MP3, WAV, M4A, OGG, FLAC or AAC file with at least three seconds of clear speech and you get a four-step verdict (AI Generated, Likely AI, Likely Human, Human) plus an AI probability score. Longer clips are split into segments, so you can see which part of a recording looks synthetic, which matters when someone splices a few fake sentences into real audio. You can check without an account; a free account saves results and share links. The site says its free tier runs the same model as paid plans, which add volume, longer clips (up to 10 minutes), exports and an API. It tells you whether a voice is likely AI, not which tool made it. See our AI Voice Detector tool page.

Note: the pricing tables of AI Voice Detector and Deepfake Detector (deepfakedetector.ai) are identical, down to the yearly totals. We did not confirm whether the two sites share an owner or a model, so don't count them as two independent opinions.

ElevenLabs AI Speech Classifier

ElevenLabs is one of the most widely used voice generators, so if you suspect a clip was made with it, this is the most precise free check: it returns the probability that the audio came from ElevenLabs and shows similar voices from its library. It uses only the first minute and needs no login. Two limits matter. It does not detect other generators, so a low score does not mean "human". And ElevenLabs says it does not reliably classify audio from its own Eleven v3 model, which is the expressive model people most often use for realistic fakes. More on the platform on our ElevenLabs page.

Gemini verification (SynthID)

Upload an audio file at gemini.google.com and ask "Was this created or edited by Google AI?". Gemini looks for SynthID, an inaudible watermark Google adds to audio from its own models, and for Content Credentials. A positive result is strong evidence. A negative one only means "not made by Google AI": Google states that Gemini currently recognizes only content from Google's tools, even though other companies have started adopting SynthID.

Hiya and Resemble AI browser extensions

Both check audio and video while you browse, which saves downloading a clip from social media first. Hiya's Deepfake Voice Detector is a free Chrome extension focused on voices. Resemble AI's extension is in beta, needs a free Resemble account and gives 4 scans a day; it also covers images and video. Resemble mainly sells detection, watermarking and voice-identity tools to businesses. See our Resemble AI tool page.

For journalists and researchers: DeepFake-o-Meter

The University at Buffalo's DeepFake-o-Meter runs a file through several independent research detection models and shows their results side by side. It requires signing in and is slower and more technical than the consumer tools, but seeing several models disagree is a useful reality check.

How to check a recording, step by step

  1. Get the best copy you can. Download the original file or voice note. A screen recording or a clip re-recorded from speakers loses exactly the details detectors look for.
  2. Cut it to the speech. Trim intros, music and long silences. Keep at least several seconds of one person talking clearly.
  3. Run two different kinds of checks. One general detector (AI Voice Detector, Resemble or Hiya) and one provenance check (the ElevenLabs classifier if you suspect it, Gemini for SynthID). They answer different questions.
  4. Read segment scores, not just the headline. A real interview with a few inserted fake sentences may score "human" overall but spike in one segment.
  5. Check the context. Who posted it first? Does the original source, the speaker's own channel or a reputable outlet have the same clip? Is there a longer version? Context debunks more fakes than audio analysis does.
  6. Write down what you checked. Tool, date, score and a link to the result. If you report the clip to a platform, an employer or the police, this is what they will ask for.

How to read the result

  • "Likely AI" from two unrelated tools plus a suspicious context: treat the clip as fake until proven otherwise.
  • A watermark or Content Credentials saying AI was used: strong evidence, because it is a positive signal from the generator, not a guess.
  • "Human" from one tool: weak evidence. New or unknown generators, phone-line quality and compression can all hide AI speech.
  • Tools disagree: common, and an honest answer. Fall back on context and on verifying with the person directly.

Warning signs you can hear (and why they are no longer enough)

In 2024 the advice was to listen for robotic tone, odd pauses and flat emotion. Current models have mostly fixed those, and real people on bad phone lines can sound robotic too. Some cues still show up, especially in cheaper tools and live voice conversion:

  • Breathing that never happens, or happens at perfectly regular intervals.
  • Odd stress on names, numbers, addresses or words in another language.
  • Emotion that doesn't match the words, or the same intensity for a whole minute.
  • Background noise that cuts in and out exactly with the speech, or a "studio" voice on a supposedly noisy street.
  • In a live call, a short delay before every answer and trouble with interruptions or follow-up questions.

Use these as a reason to check further, never as proof either way.

Rules that help you spot AI voices

  • EU: from 2 August 2026, Article 50 of the AI Act requires providers of generative AI to mark their output so it can be detected, and those who publish deepfake audio, images or video to label them. A voluntary EU Code of Practice describes how to do it.
  • YouTube: creators must disclose realistic AI-generated or altered content, including making a real person appear to say something they didn't, and YouTube labels it.
  • US phone calls: AI-generated voices in robocalls fall under the TCPA's consent rules (FCC ruling, 8 February 2024).

Labels and watermarks only help when the person making the fake follows the rules. Scammers don't, so the absence of a label proves nothing.

Common mistakes

  • Trusting a single score. Every detector misses some AI audio and flags some human audio.
  • Treating a generator-specific "no" as "human". The ElevenLabs classifier can only say whether ElevenLabs made it.
  • Checking a degraded copy. Screen recordings and re-uploads strip the signal; find the original file.
  • Arguing with a caller about whether they're real. Hang up and call back. It takes thirty seconds and doesn't depend on any technology.
  • Posting your voice everywhere without a plan. Voice-cloning tools need only a short sample, and public videos are an easy source. If you or your family post videos, agree on a code word now.

Curious how realistic synthetic voices have become? Browse the AI voice generators we track or our guide to voice cloning software.

FAQ

Is there a free AI voice detector?

Yes. AI Voice Detector (aivoicedetector.com) gives 50 free checks a month for clips up to two minutes. The ElevenLabs AI Speech Classifier is free without an account but only recognizes ElevenLabs audio. Gemini can check audio for Google's SynthID watermark. Hiya and Resemble AI offer free Chrome extensions for audio and video you play in the browser.

Can AI voice detectors be wrong?

Yes, in both directions. Compression, background music, phone-call quality and re-recording a clip from speakers all make results less reliable, and new voice models appear faster than detectors are retrained. ElevenLabs itself says its classifier does not reliably recognize audio from its own Eleven v3 model. Treat a score as one piece of evidence, not a verdict.

Can voice cloning be detected?

Often, but not always. A cloned voice is still synthetic speech, so general detectors can flag it, and a generator-specific tool can recognize audio from its own platform. A short, noisy or heavily processed clip may not carry enough signal. For a call that asks for money or codes, verifying the person through another channel is more reliable than any detector.

Can YouTube detect AI voices?

YouTube does not publish an AI-voice detector for viewers. It requires creators to disclose realistic AI-generated or altered content, including content that makes a real person appear to say something they didn't, and labels those uploads. A missing label does not prove a voice is real.

What happened to the PlayHT AI voice classifier?

PlayHT is no longer available, so its voice classifier, which our 2024 version of this guide recommended, cannot be used. If you are looking for a free replacement, start with AI Voice Detector for a general check and the ElevenLabs classifier if you suspect ElevenLabs.

How do I protect my family from AI voice-clone scams?

Agree on a family code word that is never posted online, and make a rule that nobody sends money or codes because of a call or voice message alone. If a relative calls in distress, hang up and call them back on the number you already have. Report attempts in the US at ic3.gov.