Recent updates

AI Song Checker V4.0

Like it or not, the AI music space keeps "moving forward," with more players jumping into the scrum every month or two. That means I've gotta roll out regular updates. So, here's version 4.0!

Update September 16, 2026:

Version 4.3 now uses stem splitting for hybrid results that score >= 85%. This means each result takes a bit longer, but by analyzing vocals and instrumental separately, we can tell whether the >= 85% score is appropriate. If the instrumental comes back human while the vocals comes back AI, the final result is adjusted down to better reflect the human element, and vice versa.

Latest models supported:

Our core model still supports Suno, Udio, and Riffusion. Since our last update, six decent new models have emerged:

  • ElevenLabs 2.0
  • MiniMax Music
  • Mureka v8.0
  • Treblo v3.0 (formerly Sonauto)
  • Lyria 3 (Google)
  • FlowMusic (Google)

Each of these gave the old detector a run for its money. The new v4.0 SubmitHub model spots them easily.

Treblo has also provided an open-source model that we're using for secondary verification.

Methodology:

My methodology for this version was pretty much the same as v3: source a few hundred (or thousand) examples from each platform, then spin out different quality tiers (eg, 128kbps, 192kbps, WAV, mp4, flac).

From there, I run two analyses:

  • Spectral: basically, turning each song into a "picture" that can be analyzed and compared
  • Temporal: the actual features/characteristics of a song, such as tempo, phase and timing alignment, loudness, and more

You can see a breakdown of both scores when running an analysis. Sometimes they disagree. When they both agree, it's a very strong signal.

Accuracy:

One of the things that I did differently here is to set up a "hold out" group to test on - a set of songs that the training models never actually see. I know, it's a bit shocking this is the first time I've done this. A hold out group is fairly standard for this kind of testing, but somehow I managed. What I can tell you is that based on this results, this model is pretty rock solid.

  • Human detection: got 499/500 predictions right (only one false positive, an ASMR track that came back as ~55% AI)
  • All the other models: 99%+ accuracy

So, can you trick the detector? Yep, you sure can. It's pretty much 100% accurate when looking at a song downloaded directly from one of the AI platforms. But as soon as you start fiddling with stems, production, and more, it gets a lot muddier.

What's next?

Oh, um, not much. I've been working on an upgraded genre detector for 2 months now. Pulling my hair out, at this point, which isn't a good thing to do as a middle-aged man. This AI detector model is pretty buttoned up, so I don't anticipate any major changes until the next big AI music model comes along (Suno 6???).

Want API access? If you work for a label or distribution company, reach out via SH Labs. We're already powering the AI detection for a few cool platforms, including Traxsource, Bandcamp, GEMA, and a number of other distributors.