In this paper we propose a new representation as input of a Convolutional Neural Network in the goal of detecting music structure boundaries. For this task, previous works used a late-fusion of a Mel-scaled Log-Magnitude Spectrograms (MLS) and a lag matrices networks. We propose here to use several self-similarity-matrices, each representing different audio descriptors, and combined using the depth of the input layer. We show that this representation improve the results over the use of the lag-matrix. We also show that using the depth of the input layer provide a convenient way for early fusion of representations.
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 and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.