Detection of repeating items in audio streams using data-driven ALISP sequencing

Radio streams often contain redundant parts. Commercials on radio or television stations, songs on music channels and jingles broadcasted before a specific radio or TV show, are some of the repeating objects in multimedia streams. In this paper, an audio fingerprinting system to detect repeating objects in audio streams is proposed. In order to resolve this problem, the ARGOS segmentation framework is used. This framework is combined with the ALISP-based audio fingerprinting system to build a new audio motif detection system. An approximate string matching algorithm inspired from BLAST technique is applied to speed up the approximate string matching to find the repeating items in the audio streams. Most of the audio motif discovery systems proposed in the literature are evaluated on repeating songs with long duration (about 5min). In our case, the ALISP-based system is evaluated on advertisements and songs where the duration could vary from few seconds to some minutes. The system is evaluated on 21 days from 3 French radio stations. On a set of 3081 repeating songs and 1315 repeating advertisements a mean recall rate of 98% with the corresponding precision value of 99% were achieved. The results show that the system is robust against different kinds of distortions present in radio streams.