Abstract:
This paper presents a methodology for speech or music
classification over broadcast digital audio signals. In feature
extraction, only a single feature named spectral entropy is
employed. The spectral features of recorded audio stream are
processed by means of shorter basis. Classification framework
is based on an efficient region growing technique that bears its
origins in the field of image segmentation. The efficiency of this
classification approach is investigated over a range of real
audio streams and generated data sets consisting of news from
internet radio stations such as BBC, United Nations and music
experts extracted from CDs. These datasets include a number of
male and female speakers and music clips of different genres:
pop, classic and rock. The results indicate that this system can
be used in segmentation and classification of speech and music
audio data with accuracy reaches 97%. The system is
implemented with Mat lab programming language.
