Publication:

Detecção automática de música

This document presents the work done in automatic detection of music events present in audio signals. The signal corresponds to recordings of broadcast radio and TV programs. The aim of this work is the development of algorithms to discriminate musical segments from other sounds. The algorithms require the definition of models for audio classification, in our case Gaussian models. Two models were created, one for music and another for non-music (a background model representing speech in most of the times) and the classification is based on log likelihood ratios. In the first part of this work several hours of audio recordings have been manually annotated in order to define an audio database. The database includes two sets of audio files, one set for model training and another to test the detection system. The proposed method, despite its simplicity, has proven capable of achieving good results. Ultimately the intent of this project is to construct a series of algorithms to differentiate, in the best way possible, different audio events that are present in the files, including silence, music, speech and other events.