Free tool
Spectrogram viewer
Turn a track into a picture of its frequency content, then read it honestly: where the codec cut the top of the band, how much energy sits in the air region, and why none of that identifies who made the music. Decoding and analysis happen in your browser.
Short answer
Can you tell a song is AI-generated by looking at its spectrogram?
No. A spectrogram reveals encoding and mixing history — lossy cutoffs, band balance, dynamics — and those look the same whether a human or a model produced the music. It is a genuinely useful way to understand what happened to a file, and it is not a detector.
How detection actually works →Drop an audio file here
MP3, WAV, FLAC, M4A, Ogg or Opus. Up to thirty seconds from the middle is drawn.
What people expect a spectrogram to show, and what it does
The most common request we see is a visual tell: a shape, a band, a smear that marks a track as machine-made. It does not exist in any form that survives normal distribution. Everything visible at this resolution is the product of the mix, the master and the codec, and generated tracks pass through the same three stages as everything else. The lowpass shelf people circle as proof of AI is a bitrate decision made by an encoder.
What the picture is good for is context. It tells you whether you are looking at a lossless master or a third-generation re-encode, whether the top end was rolled off, and whether the loudness sits flat across the excerpt. All of that changes how much weight any acoustic estimate deserves — we measured how much on the compression study.
Reading the axes
Time runs along the horizontal axis, frequency up the vertical, and brightness maps level in decibels over a ninety-decibel window. Horizontal lines are sustained tones. Vertical stripes are transients: drums, plosives, edits. A dark band across the top with a clean boundary is a codec cutoff. A gradual fade to dark at the top is a mixing or mastering choice.
If you want the measured version of these observations rather than the visual one, the detector reports the same underlying features as numbers, with an explanation of each and the cases where they mislead.
Use it alongside the other free checks
Provenance first: read the file's tags, since a written claim beats any inference. Then look at the spectrogram to understand what has been done to the audio. Then, if the question is still open, take a probability from the detector and treat it as one piece of evidence rather than a verdict.
For a full walkthrough of the shapes, the codec myths and the findings that hold up, read how to read a spectrogram.
Questions about spectrograms
- Can a spectrogram tell you a song is AI-generated?
- No. A spectrogram shows how energy is distributed across frequency and time, which is dominated by the mix, the mastering and the encoder — all of which are identical for human and generated material. Claims that a single visual artefact identifies AI music do not survive controlled testing.
- What is the flat line near 16 kHz on my spectrogram?
- That is a lossy codec cutoff. MP3 and AAC encoders discard the top of the band to save bits, leaving a hard horizontal edge, usually between 15 and 20 kHz depending on bitrate. It tells you the file was compressed, which is true of almost everything distributed online.
- Is my audio uploaded to a server?
- No. The file is decoded by your own browser and the transform runs in the page, exactly like the detector on this site. Nothing is transmitted, stored or logged.
- Which part of the song is shown?
- Up to thirty seconds taken from the middle of the file, because intros and fades are the least representative parts of a track. Longer files are excerpted rather than downsampled so the detail stays readable.
- Why does my lossless file look different from the same song on a streaming platform?
- Because the platform re-encoded it. Transcoding changes the top of the band and can alter fine detail across the picture, which is one reason a visual comparison between two versions of the same recording is rarely conclusive.