Dual-engine approach for detecting AI-generated music in compressed audio

The author explores a hybrid approach to detect AI-generated music, combining a CNN-based model and a source separation engine, to overcome the limitations of CNN-only models on compressed audio formats like MP3.

Details

The author was working on detecting AI-generated music and faced the same issue as Deezer's team - CNN-based models trained on mel-spectrograms perform well on uncompressed WAV files but break down when the audio is compressed to MP3. To address this, the author added a second engine based on source separation using the Demucs model. The idea is to separate the audio into 4 stems (vocals, drums, bass, other), remix them, and measure the difference between the original and reconstructed audio. For human-recorded music, the stems bleed into each other during recording, so the reconstruction process produces noticeable differences. For AI-generated music, where each stem is synthesized independently, the reconstruction yields nearly identical results. This hybrid approach achieved a human false positive rate of ~1.1% and an AI detection rate of 80%+, working across different audio codecs. The limitations include varying detection rates across different AI generators, non-deterministic behavior of Demucs in borderline cases, and only being tested on music (not speech or sound effects).