Content based event retrieval on TV broadcast audio

Auditory data contains important information about the content of multimedia data. This paper presents a method for content based event retrieval on broadcast audio. The aim of this study is to retrieve audio events from huge multimedia databases. 17 classes which are most frequently observed in TV broadcast, and which are considered as an important input to higher level semantic analysis of multimedia data are selected. Audio streams are divided into homogenous segments in order to generate fingerprints that describe both temporal and spectral information of audio events. Both spectral and temporal properties of audio events are analyzed and some fingerprints to represent these properties are presented. Audio events are modeled by Gaussian Mixture Models. For the retrieval, an ordered sequence is provided to the user for each event, sorted by the likelihood values of the fingerprints. The system aims to bring the query events with higher likelihood values first. Mean average precision value is used to evaluate retrieval performance.17 audio classes are tested on 11 hours of TV recordings and 18,5% average precision is achieved.