Speech Music Discrimination in mixed Audio Content

Speech Music Discrimination (SMD) can highly improve Automatic Speech Recognition (ASR) and Music Analysis results of broadcast or cinema audio signals. Especially in infotainment productions or commercials, speech is often embedded in background music. We present an audio segmentation system for television commercials, optimized for the handling of speechmusic mixture classes using independentt detectors for the classes "speech", "music", "other" (sounds) and "silence". The system is optimized to process high and low quality audio in terms of bandwidth limitation and distortion. We evaluated the system on a representative set of commercial spots in three different languages with different audio qualities.