Content area
This paper addresses the issue of distinguishing commercially played songs from non-music audio in radio broadcasts, where automatic song identification systems are commonly employed for reporting purposes. Service call costs increase because these systems need to remain continuously active, even when music is not being broadcast. Our solution serves as a preliminary filter to determine whether an audio segment constitutes “music” and thus warrants a subsequent service call to an identifier. We collected 139 h of non-consecutive 5 s audio samples from various radio broadcasts, labeling segments from talk shows or advertisements as “non-music”. We implemented multiple data augmentation strategies, including FM-like pre-processing, trained a custom Convolutional Neural Network, and then built a live inference platform capable of continuously monitoring web radio streams. This platform was validated using 1360 newly collected audio samples, evaluating performance on both 5 s chunks and 15 s buffers. The system demonstrated consistently high performance on previously unseen stations, achieving an average accuracy of 96% and a maximum of 98.23%. The intensive pre-processing contributed to these performances with the benefit of making the system inherently suitable for FM radio. This solution has been incorporated into a commercial product currently utilized by Italian clients for royalty calculation and reporting purposes.
