Voxalized

How to Tell if a Song Is AI: What You Can Hear and What You Can Measure

Most listeners cannot tell an AI song from a human one. Here is what a producer listens for, which popular tips do not hold up when you measure them, and how a detector finds what ears miss.

Kilian8 min read

Most people cannot tell an AI song from a human one. In a Deezer and Ipsos survey of 9,000 listeners in eight countries, 97% could not reliably pick out the fully AI-generated tracks. As a producer I find it one of the easier things to hear, and almost always it is the vocal that gives a song away. This guide covers what I listen for, which popular tips do not hold up when you measure them, and what an AI music detector reads in a file that no ear can.

The quick check

What to check What points to AI How much it tells you
Vocal top end Shiny, thin, slightly metallic air above the voice A lot
Vocal tone Nasal, the same color on every word A lot
Dynamics Everything equally loud, no quiet syllables Some
Lyrics Stock images, rhymes that fit but say nothing Some
Artist profile Many releases in weeks, no shows, no history Some
Tempo Drifting or perfectly steady Nothing, see below
High end of the file Treble stops at 16 kHz Nothing, see below
Spectrum of the file Evenly spaced peaks A lot, needs a tool

Listen to the vocal first

Instruments are hard to judge. A synth made by a model and a synth from a preset sound much the same. The voice is different, because everyone has heard human voices all their life.

The top end. AI vocals carry a bright layer on top of the voice, a kind of permanent air that sits there whether the singer is loud or quiet. On a recorded vocal the air changes with every breath and every consonant. On a generated one it is constant, and on good headphones it has a slightly metallic edge.

The nasal tone. Many generated voices sound as if they were sung through the nose. It is subtle on one line and obvious over a whole song, because the tone never changes. A real singer opens up on long notes, gets rough when pushing and thin when holding back.

The compression. Among producers this is the standard comment: it sounds overcompressed. Every syllable is as loud as the next, breaths are either missing or sit at the same level as the words, and nothing ever jumps out. A human take that was compressed this hard would pump and distort. The generated one stays smooth, which is exactly what makes it sound flat.

Listen also for consonants that smear into each other, words that are almost a word, and doubled or harmony voices that follow the lead a little too exactly.

Read the lyrics

Generators write the average of every lyric they were trained on, so the same images come back again and again. "Neon lights" is the one I notice most, above all in deep house. City nights, shadows, echoes, whispers and fire in the veins are close behind. The lines rhyme and scan, and at the end of the song you could not say what it was about.

One stock phrase proves nothing, since human writers use them too. A whole song of them, sung by a voice with the signs above, is a pattern.

Look at the artist

A real act leaves traces: live dates, photos from more than one shoot, older releases that sound less finished, other people who mention them. Signs worth a second look are a catalogue that appeared within a few weeks, dozens of releases a year across unrelated genres, cover art in one generated style, and no person behind the name anywhere.

Streaming services have started to help. Spotify marks profiles with a generated identity, which we explain in Spotify's AI Persona label, and Deezer tags fully generated tracks.

When we built our detector we tested the tips that circulate, on Suno songs and on finished masters and vocal recordings from our own catalogue. Two of them failed.

Tempo. The claim is that AI songs drift off the grid, or are suspiciously steady. On the files we measured, the human recordings drifted more than the Suno songs. Tempo separated nothing.

The missing high end. Generated songs often stop somewhere between 16 and 20 kHz. So does every MP3, and so does everything you hear on a streaming service. A cutoff shows how a file was encoded. It says nothing about who made it.

Both were dropped from the detector. They are worth knowing mainly so that nobody is called out over an MP3.

What a detector measures

A generator does not record sound, it computes it. In the last step a network turns a compressed description of the song into a waveform by upsampling it several times. Researchers at Deezer showed in their paper A Fourier Explanation of AI-music Artifacts that this step leaves small, regular peaks in the spectrum, and that the peaks come from the architecture of the model, whatever music it makes. They are far too quiet to hear.

Our AI music detector averages the spectrum of a whole song and looks for that pattern. It runs in your browser, so the file is never uploaded. It takes three readings:

  1. The fingerprint. A model trained on tens of thousands of Suno and Udio songs and human recordings compares the peaks between 1 and 8 kHz.
  2. The frequency grid. A second, independent measurement checks whether the treble is louder on exact steps of 200 Hz.
  3. The file tag. Suno writes its name into the files it exports. If the tag is still there, the question is settled. A separate guide explains what is in a Suno file.

A common trick is to speed a song up or slow it down by a few percent, which moves the pattern. The detector reads the spectrum at 1,600 playback speeds and finds it again.

Where detection stops

No detector is proof. On our own test files, every untouched Suno song was recognized and none of the finished human masters was flagged. Heavy processing weakens the trace, one pitch-shifted version slipped through, and generators other than Suno and Udio are not well covered yet.

A song can also be partly generated: a human production with an AI vocal, or an AI demo that was re-recorded by a singer. A detector reads the file, not the story behind it. Treat a result as one more sign next to what you hear, and do not accuse anyone because of a percentage.

Why it matters before you release

For me as a listener it is simple. When I notice that a song is generated, I skip it or hide it from the playlist. I cannot connect with it at all.

For producers the question is more practical. AI vocals are used commercially today, and for now that is tolerated, but the legal ground is not settled and the big generators are still being sued by major labels. We go through that in detail in whether AI vocals are royalty-free. In sync, meaning ads, trailers and film, you do not need to show up with a generated vocal at all. Clients there check very carefully where every part of a track comes from, and I expect that to get stricter.

AI vocals sold as real ones

This is the part that annoys me most. A growing number of sites advertise huge vocal catalogues that are entirely generated, and still charge around 30€ per vocal. In my view that is a rip-off. The price of a vocal pays for a singer who sits down, records, and delivers a mixed and mastered take. A generated vocal costs the seller next to nothing, and what you get is a generic voice that many other buyers have in almost the same form. It will not help you build an audience.

Beginners fall for it most easily. A generator produces the average of everything, and the average sounds good on first listen. Many people buy a pack, build a track around it and only notice months later what they bought. Before you buy vocals anywhere, check three things: is the singer named, does that person exist outside the shop, and does the license say who recorded the vocal? Our guide to Splice alternatives covers what to check before you buy a vocal pack.

On Voxalized every vocal comes from a named singer with a profile, and we review every singer before a vocal goes on sale.

This article reflects our own measurements on a small set of files and my experience as a producer. It is not legal advice.

Frequently asked questions

  • How can you tell if a song is AI generated?

    Listen to the vocal first: a thin, shiny top end, a nasal tone and heavy compression are the most common signs. Then read the lyrics for stock phrases and look at the artist profile. For a second opinion, run the file through an AI music detector, which reads a pattern in the spectrum that you cannot hear.

  • Can people hear the difference between AI music and human music?

    Most cannot. In a Deezer and Ipsos survey of 9,000 people in eight countries, published in November 2025, 97% could not reliably tell fully AI-generated songs from human ones. Trained ears do better, mainly on the vocal.

  • Do AI songs have a tempo that drifts?

    Not in a way that helps. We measured tempo drift on Suno songs and on songs from our own catalogue, and the human recordings drifted more than the generated ones. Tempo is not a reliable sign.

  • Does a missing high end mean a song is AI?

    No. Generators often cut the treble somewhere between 16 and 20 kHz, but so does every MP3 and every stream. A treble cutoff tells you how a file was encoded, not who made it.

  • How accurate are AI music detectors?

    Good on untouched Suno and Udio files, weaker on songs that were heavily processed or made with newer generators. A result is a probability, not proof, and it should never be used alone to accuse an artist.

Summarize with AI

Share

More insights

Keep all royalties and release your song with our vocals.

Every license is personalized to you. No splits, no clearances.