Automatic removal of music tracks from tv programmes

Abstract

This work pertains to in the research area of sound source separation. It deals with the problem of automatically removing musical segments from TV programmes. The dissertation proposes the utilisation of a pre-existant music recording, easily obtainable from o cially published CDs related to the audiovisual piece, as a reference for the undesired signal. The method is able to automatically detect small segments of the speci c musictrack spread among the whole audio signal of the programme, even if they appear with time-variable gain, or after having su ered linear distortions, such as being processed by equalization lters, or non-linear distortions, such as dynamic range compression. The project developed a quick-search algorithm using audio ngerprint techniques and hash-token data types to lower the algorithm complexity. The work also proposes the utilisation of a Wiener ltering technique to estimate potential equalization lter coe cients and uses a template matching algorithm to estimate time-variable gains to properly scale the musical segments to the correct amplitude they appear in the mixture. The key components of the separation system are presented, and a detailed description of all the algorithms involved is reported. Simulations with arti cial and real TV programme soundtracks are analysed and considerations about new future works are made. Furthermore, given the unique nature of this project, it is possible to say the dissertation is pioneer in the subject, becoming an ideal source of reference for other researchers that want to work in the area.