Audio Watermark Detector

What this audio watermark detector actually checks

An AI audio file can carry two kinds of marks. The first is metadata: tags written into the file container, such as ID3 in MP3, RIFF INFO chunks in WAV, Vorbis comments in FLAC and OGG, iTunes-style tags in M4A, and provenance records like C2PA Content Credentials or the XMP field digitalSourceType. The second is an inaudible watermark mixed into the sound itself, designed to survive re-encoding. This tool reads the first kind. It does not decode the second, and nobody outside the company that embeds such a mark can, unless the company publishes a detector.

Here is what happens when you upload a file:

In those areas it looks for C2PA and JUMBF markers, XMP provenance fields and generator names: Suno, Udio, ElevenLabs and others from the same list it uses for images and video. Each hit becomes a row in the findings table with a risk level: generator names and C2PA count as high, the spectral anomaly as medium.

Suno watermark detector: what you can and cannot see

Suno has said it uses inaudible watermarking to recognise its own tracks, but it has not published a public checker for that mark. So no third-party Suno watermark checker can confirm it, including this one. What the tool can show is whether the file still carries a Suno reference in its tags or metadata. Files downloaded straight from the service are the best case. A track that was exported from a DAW, re-encoded by a distributor or ripped from a streaming platform will usually have new tags, and the reference disappears.

How to read the result:

• "Generator signature: Suno (audio)" means the string suno appeared in the scanned metadata. The report does not say which tag held it; to see the field, open the file in a tag editor or run ffprobe on it.
• "No clear watermarks found" means the metadata is clean. It does not mean the track is human-made.
• A short name can match inside an ordinary word. "Udio" is part of "Studio" and "Audio", so an encoder tag like "FL Studio" or "Audio Studio" can raise a false Udio hit. Look at the tags yourself before drawing conclusions.

Udio and ElevenLabs

Udio does not publish a watermark checker either; the same metadata rules apply. ElevenLabs is different: it offers its own AI Speech Classifier, which estimates whether a clip was generated with ElevenLabs. If you suspect an ElevenLabs voice, run the clip through that classifier, then use this tool to see what the file's metadata says. The two answer different questions: one analyses the sound, the other the container.

The high-band check on WAV files

For 16-bit PCM WAV, the tool takes up to the first 20 seconds of the first channel, computes the spectrum and measures how much energy sits in the top 10% of the band, for example above about 19.8 kHz at a 44.1 kHz sample rate. Natural speech and most music have little there. If the share is over 2%, the report shows "Unusual high-frequency energy". That is a hint worth a closer look in a spectrogram, not proof: bright cymbals, synth noise, dithering and some encoders also push energy there, and a watermark can just as well sit in the audible band where this check does not look. 24-bit and 32-bit WAV files skip this check.

AI music detector: how to judge "is this song AI"

No single test answers this. Combine the file check with manual signs:

1. Check the file you received, not a re-upload. The original download keeps tags; a YouTube rip or a phone recording keeps nothing.
2. Look at the distribution data. Deezer labels albums its own detector considers fully AI-generated. Credits that list no performers, producers or writers are a soft signal.
3. Listen for typical artefacts. A metallic "shimmer" on cymbals and sibilants, consonants that smear, lyrics that drift into nonsense, instruments that change timbre mid-phrase, a mix that sounds the same in every section.
4. Look at the spectrogram. Any audio editor can show one. A hard cut-off at the same frequency throughout, or bands of noise that do not follow the music, deserve questions.
5. Ask for stems or project files. A human producer can usually supply them; for client or label work this is the most reliable check.

AI song detector: how to read each result

The report lists every signal it found. Here is what each one means when the question is "was this song made with AI":

Suno, Udio and ElevenLabs at a glance

Suno watermark remover: what cleanup does and does not do

For WAV, the cleaned copy keeps only the format, audio data and fact chunks. All tags, text chunks and embedded provenance records go, and the sound is not touched. For MP3, FLAC, OGG, M4A and AAC the report gives an ffmpeg command instead:

ffmpeg -i in.mp3 -map_metadata -1 -c copy out.mp3

It copies the audio stream without re-encoding and drops the global metadata. Note that ffmpeg writes its own encoder tag by default; that is normal.

What no metadata cleanup can do is remove an inaudible watermark mixed into the audio. Re-encoding, pitch shifting or adding noise may weaken such a mark or may not; the tool makes no claim either way. Cleaning tags also does not change licensing. Suno's terms, for example, tie commercial use to a paid plan at the time of creation; stripping a reference from the file does not affect who owns the track. Use cleanup for your own files, for privacy and for tidy delivery, not to hide where audio came from.

Related checks

Video soundtracks are checked as part of the container in the video watermark detector. For cover art and thumbnails generated with AI, use the image watermark detector. Text from AI chats can carry invisible characters; the text watermark detector finds them.