Recent updates

AI Song Checker V3.0

Honestly, things are moving so fast here that it's quite hard to keep up. I'm gonna keep trying, though!

This latest version of the AI Song Checker is a pretty big overhaul of both detection models, resulting in a massive jump in accuracy and a much deeper understanding of what makes AI music sound... well, AI.

Oh, and special thanks for those of you who've been testing it out and leaving feedback. Your input helps make things better, so if the result is wrong, please use the up/down thumbs as shown below πŸ‘ πŸ‘Ž

Temporal vs. Spectral

Before we dive into what's new, here’s a super simple breakdown of the two different techniques being used:

  • The Spectral Model: This model turns an audio file into a spectrogram – a detailed picture of all the frequencies in the song. It then uses advanced image recognition to spot patterns and textures that are tell-tale signs of AI generation.
  • The Temporal Model: This model analyzes how the music unfolds over time. It listens for how melodies, harmonies, and rhythms connect from one moment to the next, checking if the song's "story" makes sense or if it has the quirky flow that AI can produce.

By combining both of their findings, I'm able to get a much more reliable and nuanced verdict.

What's Changed in Version 3?

This update focused on making both models smarter, faster, and more attuned to the latest generation of AI music tools.

Temporal Model

  • Support for Suno 4.5(+)
  • A few new "features" based on Deezer's latest research paper (this ended up contributing a ~5% improvement)
  • A bunch of optimizations to the underlying code, which makes the training and analysis process over 20% faster

Spectral Model

  • Support for Suno 4.5(+)
  • Looks at both the core sound (log-mel spectrogram) and its changes over time (deltas), giving it a much richer "picture" to analyze
  • Analyzes multiple 10-second "snapshots" of a song and averages the results
    • This makes it far more robust and less likely to be misled by a single unusual moment in the track. It also means it's better at handling short Spotify samples

The Results

Comparing the old models to the new models took a bit of time to get working, but I'm pretty happy with the story it tells.

Of course, my percentages here are based on the data I have available. I have no doubt that some of you - especially those heavily editing AI songs - will be able to fool it!

Temporal Model

I've rolled out a few tweaks to the temporal model (most-recently in June). The biggest story here is the model's newfound ability to handle the latest Suno updates. While it was great at identifying older AI, it was often fooled by Suno 4.5(+).

MetricV2.0V3.0ImprovementOverall Accuracy91.6%98.5%+6.9%Suno 4.5+ Accuracy33.0%98.0%+65.0%Execution Time2,063s1,588s+23.1% Faster

The massive 65% improvement on Suno 4.5+ tracks was the primary goal of this update, moving it from unreliable to highly accurate.

Spectral Model

The Spectral model fell behind big time (last updated in ~Feb), so these updates resulted in the most dramatic transformation.

MetricV2.0V3.0ImprovementOverall Accuracy66.5%98.6%+32.2%Modified AI Accuracy83.8%99.4%+15.6%Suno 4.5+ Accuracy26.0%99.0%+73.0%

A 32% jump in overall accuracy is huge, driven by massive gains in previously difficult categories like Riffusion and Suno. The trade-off is that the new analysis is more thorough and therefore slower (by ~1s), but the accuracy speaks for itself.

What's Next?

While V3 is a huge step up, the work is never done! The world of AI music will continue to evolve, and so will this tool. The immediate focus is on continuing to expand the training dataset with even more diverse examples, especially as new AI music models are released.

As always, I'm relying on your feedback. The rating option on each result is invaluable for identifying where the models still get it wrong. Every thumbs-up or thumbs-down helps train the next version.

Final Thoughts

This remains a passion project for me, and I'm committed to keeping it free and accessible. The goal isn't to create a perfect, infallible gatekeeper, but to build a useful tool that adds transparency to the music ecosystem. Your continued feedback, ideas, and suggestions are what make this project possible. Thank you for being a part of it πŸ’œ πŸ’œ πŸ’œ