Comparison of Effectiveness of Musical Sound Separation Algorithms Employing Neural Networks
In this paper several algorithms are presented, developed for musical sound separation. The proposed techniques for the decomposition of mixed sounds are based on the assumption that pitch of the sounds contained in the mix is known, i.e. inputs of the algorithms are pitch tracks of the signals contained in the mixture. The estimation process of phase and amplitude contours representing harmonic components is based on the limited number of inner product operations, performed on the signal with the use of complex exponentials matching pitch characteristics of the separated signals, and not on the discrete spectral representations calculated via DFT. In this paper examples of separation results are presented and each algorithm performance is analyzed. The effectiveness of separation algorithms consists in calculation of feature vectors (FVs) derived from musical sounds after the separation process is performed, and then in feeding them the Neural Network (NN) for automatic musical sound identification. The experimental results are shown and discussed. A comparison of effectiveness of all presented algorithms is also included, and conclusions are derived.
Click to purchase paper 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 $20 for non-members, $5 for AES members and is free for E-Library subscribers.