Classifying audio of movies by a multi-expert system

The paper presents a system for the automatic MPEG format. In contrast to the approaches proposed up to now, it employs a multi-expert classification system arranged according to a multi-stage architecture. The system is able to recognize not only four pure classes (music, speech, silence and noise) but also confused audio signals, such as the ones resulting from the overlap of pure audio components (for example, speech overlapped with music or noise, etc.). An extensive experimental analysis has been carried on a large audio database extracted from about 30 moving pictures recorded on low-quality magnetic media. Results confirm the effectiveness of the approach, with an average improvement of about 45% with respect to single classifier solutions.