GEOMETRY IN SOUND: A SPEECH/MUSIC AUDIO CLASSIFIER INSPIRED BY AN IMAGE CLASSIFIER
Norman Casagrande, Douglas Eck, Balázs Kégl · 2005
In this paper we adapt a well-known image processing algorithm to the task of predicting whether an audio signal contains speech or music. We derive a frame-level discriminator that is both fast and accurate. Using a simple FFT and no built-in prior knowledge of signal structure we obtain an accuracy of 87 % on frames sampled at 20ms intervals. When we smooth the output of the classifier with the output of the previous 40 frames our forecast rate rises to 93 % on the Scheirer-Slaney [1] database. To demonstrate the efficiency and effectiveness of the model, we have implemented it as a graphical real-time plugin to the popular Winamp audio player.