Convolutional neural networks have satisfactory discriminative performances for various music-related tasks. However, the models are implemented as “black boxes” and thus their processed representations are non-transparent for manual interactions. In this paper, a hierarchical sonification framework with a musical genre modeling module and a sample-level sonification module has been implemented for aural interaction. The modeling module trains a convolutional neural network from musical signal segments with genre labels. Then the sonification module performs sample-level modification according to each convolutional layer, where lower sonification levels produce auralized pulses and higher sonification levels produce audio signals similar to the input musical signal. The usage of the proposed sonification framework is demonstrated using a musical stylistic morphing example.
(See document for exact affiliation information.)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
