How to tell if someone is using a voice changer

Here is how to tell if someone is using a voice changer. Listen for a small delay before every reply, a metallic edge on s and sh sounds, a voice that cracks when the pitch jumps, and a voice that does not fit the accent or rhythm of the person speaking. Then test it: talk over them, ask them to laugh or whisper, and call back on a number you already know. To check a recording, run it through an AI voice detector.

Voice changers come in two kinds: effects that bend a real voice, and AI tools that re-voice someone as a different person. Each leaves different traces. Here is what to listen for, how to test someone live, and what a detector can and cannot tell you.

By the AI Voice Detector Editorial Team · Guide · Updated November 2026

What kinds of voice changers are there?

The answer to "can you tell" depends on which kind you are dealing with.

Kind	What it does	What it keeps from the speaker	Main tells
Effect changer (pitch, formant, robot, filters)	Processes the real voice: shifts it higher or lower, adds a robot or radio effect	Everything except the sound of the voice: words, timing, accent, breaths	Usually audible as an effect: chipmunk, giant, hollow or robotic tone
AI voice changer (voice conversion, speech-to-speech)	Re-generates the speech in a target voice, often a clone of a real person	Timing, accent, emotion and delivery; only the voice itself changes	Delay on live calls, smeared consonants, pitch cracks, a voice that does not fit the speech

AI voice changers are the harder case. ElevenLabs describes its voice changer as turning source audio into "a different, fully cloned voice" while keeping whispers, laughs, accents and emotional cues (ElevenLabs documentation). The open-source RVC project calls itself a "voice timbre conversion / voice changer framework" and recommends as little as 10 minutes of clean speech to train a new voice (RVC project README). That is why a live caller can sound like someone they are not and still answer your questions naturally. For how the target voice is built, see what voice cloning is.

7 signs someone is using a voice changer

No single sign proves anything. Two or three together are worth acting on.

A small, constant delay before every reply. Live AI conversion takes time to process each chunk of speech. RVC reports an end-to-end latency of 170 ms in its real-time mode, and 90 ms with specialist audio drivers (RVC). Phone and internet calls add delay too, so the tell is a lag that affects this one voice on every turn and makes you talk over each other more than usual.
Metallic, lisping or watery consonants. Hissy sounds (s, sh, f, ch) are hard for the model that rebuilds the audio. Listen for a buzz, a smear or a lisp that the rest of the voice does not have.
Cracks and warbles when the pitch moves fast. AI changers track the speaker's pitch and redraw it in the new voice. RVC's own feature list highlights a pitch-extraction method added to "prevent a muted sound problem", which shows how much rides on that tracking. Sudden rises, shouts, singing and fast laughter are where it slips: the voice cracks, jumps an octave or goes briefly robotic.
Laughs, coughs and breaths that sound wrong. Non-speech sounds are the hardest material to convert. Commercial tools now claim to carry laughs and whispers across, so a clean laugh clears no one, but a laugh that turns into a buzz or vanishes mid-breath is a strong hint.
A voice that does not fit the person speaking. Conversion changes the voice and keeps the delivery. A deep adult voice with a teenager's slang and pacing, or a familiar boss's voice with a different accent and different filler words, means the voice and the speaker are two different people.
Background sound that warbles or disappears. The model processes whatever the microphone picks up, not only the voice. ElevenLabs offers a setting to strip background noise for exactly this reason. Room sound that comes out warped, or total silence between words on a call that should have some, is a sign of processing.
An effect you can hear as an effect. Simple pitch shifters move the formants (the resonances that make a voice sound like a particular body) along with the pitch, which gives the familiar chipmunk or giant sound. Even with formant correction, a shifted voice often sounds hollow or boxed in. This is the easiest kind to hear.

The newest AI changers remove many of these tells, especially in recorded audio that was converted offline rather than live. Absence of signs does not mean absence of a changer. If money, access or someone's safety depends on the answer, verify the person, not the voice.

How to tell if someone is using a voice changer on a live call

A live call gives you something a recording does not: you can make the software work harder. These tests push a changer toward the edges where it slips.

Interrupt them. Talk over a sentence and listen to how fast they react. A processing delay makes overlaps clumsy.
Ask for sounds that are not speech. A laugh, a whisper, a long "sssss", a cough. Then ask them to say something loudly.
Change the pitch range. Ask them to read a phone number in a high voice, or to sing a line of a song. Pitch tracking fails most often at the extremes.
Ask what only the real person would know. A shared memory, or a family code word agreed in advance. A changer copies a voice, not someone's knowledge.
Hang up and call back on a number you already have. This is the check the US Federal Trade Commission recommends when a voice on the phone sounds like a relative asking for money (FTC consumer alert). It works whatever the software, because the scammer does not control the real person's phone.

A skilled user with good hardware can pass the first three. The last two do not depend on audio at all, which is why they are the ones to rely on. Calls that use a cloned voice to ask for money or codes are a form of vishing (voice phishing).

How can you tell on Discord, in games or in voice chat?

In gaming and voice chat, voice changers are common and usually harmless. The same signs apply, with two cautions:

Compression causes false alarms. Low-bitrate voice chat, noise suppression and a bad microphone all make real voices sound robotic or metallic. Judge the signs that compression does not create: a constant per-person delay, pitch cracks on laughs, and a voice that does not fit the speaker.
Asking is often the fastest check. Many people use effects openly. If it matters, for example before you share personal details with someone you met online, ask them to turn it off or switch to video.

If you record a call or voice chat to check it, look up the law where you and the other person are. Some places require everyone on the call to agree to a recording.

How do you check a recording for a voice changer?

When you have the audio as a file (a voice note, a voicemail, a recorded call, a clip from a video) a detector can measure what your ears cannot. An AI voice changer does not pass your voice through; it generates new audio in the target voice. RVC, for example, credits the VITS and HiFi-GAN speech synthesis models among the components it builds on. That generated audio is what an AI voice detector is built to score. Researchers have treated converted speech and synthesized speech as the same family of attack for years: the ASVspoof 2019 benchmark was built from "synthesized, converted and replayed speech" (ASVspoof 2019, arXiv).

Save the clip as MP3, WAV, FLAC, OGG, M4A or AAC, under 10 MB.
Pick the cleanest 60 seconds, where the person talks with little music or background noise.
Upload it to the free AI voice detector. No account is needed and the free check keeps nothing.
Read the verdict and the segments. You get one of four verdicts (AI generated, likely AI, likely human, human), an AI probability score and, on longer clips, a score for each segment. A spike in one segment can mean the changer was switched on part way through.
Check a second section if the result is borderline. Agreement across sections is more convincing than one read.

For longer files and more checks, a free account allows 50 checks a month on clips up to 2 minutes. Teams checking calls at volume can use the API.

What can a voice changer check not prove?

Simple effects are outside what the detector measures. A pitch shifter or robot filter processes a real human voice; it does not generate a new one. Those are usually easier to hear than to measure.
We do not publish a separate accuracy figure for voice conversion. The detector scores whether speech is AI-generated; it does not report which method made it.
It does not name the tool or the person being imitated. A high score says the voice is likely synthetic, not whose voice it was meant to be.
Phone audio and re-recording blur the signal. A recording of a speakerphone, or a clip passed through several apps, is the hardest case. Use the original file when you can.

The short version of how to tell if someone is using a voice changer: listen for the seven signs, test the person live, verify them through a number you already trust, and check any recording with a detector.

For the full set of listening checks on AI speech in general, see how to detect AI voices.

Frequently asked questions

How can you tell if someone is using a voice changer?

Listen for a constant small delay before each reply, metallic or lisping s and sh sounds, cracks when the pitch jumps, laughs or breaths that sound wrong, and a voice that does not fit the speaker's accent or rhythm. On a live call, interrupt them, ask for a laugh or whisper, and call back on a number you already know. For a recording, run it through an AI voice detector.

Can a voice changer be detected?

Often, but not always. Simple pitch and robot effects are usually audible. AI voice changers generate new speech in the target voice, and that synthetic audio is what AI voice detectors score, so a recording can be checked. The newest tools remove many audible tells, so verify the person through a channel you trust when it matters.

Do voice changers have a delay?

Live AI voice changers do. The open-source RVC project reports about 170 ms end to end in real-time mode, and 90 ms with specialist audio drivers. Calls add their own delay, so the tell is a lag that is constant for one person on every reply.

How do you tell if someone is using a voice changer on Discord?

Listen for pitch cracks on laughs, a constant delay on that one person, and a voice that does not fit how they talk. Be careful with robotic or metallic sound alone, because voice chat compression and noise suppression cause it in real voices too. Asking them to turn the effect off, or to switch to video, is often the quickest check.

Can a voice changer make someone sound like a specific person?

Yes. AI voice changers can convert speech into a cloned voice. RVC recommends about 10 minutes of clean speech to train a voice, and commercial services can use any voice cloned in the account. That is how scammers impersonate relatives and executives on live calls.

Have a recording that sounds off? Check it free.

Check a recording