Enhanced TV Broadcast Monitoring with Source Separation-Assisted Audio Fingerprinting

Audio fingerprinting (AFP) algorithms identify music in audio recordings. Traditionally, studies have focused on increasing the robustness of AFP systems to pitch and tempo alterations. On the other hand, solutions for background music identification have been less explored even though they are crucial for ensuring proper royalty distribution in television broadcasts, for instance. In this case study, we investigate whether incorporating source separation as a preprocessing step before applying audio fingerprinting can enhance music identification. We study the impact of eight public source separation algorithms, two private ones, and different trained models based on UNet and WaveUNet. We evaluate them together with four public AFP algorithms and an additional AFP model trained for background music identification. We also frame the study with a computational cost benchmark, which helps assessing the impact of incorporating the source separation preprocessing. The results demonstrate that source separation algorithms directly improve the audio fingerprinting performance in TV broadcast monitoring, directly contributing to better royalty distribution. They also show that the enhancement is not only limited to background music but also to foreground music, making this approach an interesting solution for other music identification use cases.