Automatic detection of known advertisements in radio broadcast with data-driven ALISP transcriptions

This paper describes an audio indexing system to search for known advertisements in radio broadcast streams, using automatically acquired segmental units. These segmental units called ALISP units are acquired automatically using temporal decomposition and vector quantization and modeled by Hidden Markov Models (HMMs). To detect commercials, ALISP transcriptions of reference advertisements are compared to those of radio stream using the Leven-shtein distance. The system is described and evaluated using broadcast streams provided by YACAST. On a set of 802 advertisements we achieve a mean precision of 95% with the corresponding recall value of 97%. The results show that the system is robust in situations where the advertisement to detect is stretched or suffer from time distortions. Moreover, this system allowed us to detect some annotation errors.