Audio Identification Using Sinusoidal Modeling and Application to Jingle Detection

This article presents a new descriptor dedicated to Audio Identification (audioID), based on sinusoidal modeling. The core idea is an appropriate selection of the sinusoidal components of the signal to be detected. This new descriptor is robust against usual distortions found in audioID tasks. It has several advantages compared to classical subband-based descriptors including an increased robustness to additive noise, especially non-random noise such as additional speech, and a robust detection of short audio events. This descriptor is compared to a classical subband-based feature for a jingle detection task on broadcast radio. It is shown that the new introduced descriptor greatly improves the performance in terms of recall/precision.