AUTOMATIC CHORD ESTIMATION ON SEVENTHSBASS CHORD VOCABULARY USING DEEP NEURAL NETWORK

This paper proposes an automatic chord estimation (ACE) system with a two-layer architecture. The first layer performs chord smoothing with "GMM + HMM" approach. Then given the results of the first layer, the second layer performs chord estimation using a deep neural network, which is trained on a well chord-type balanced dataset. The system accepts exactly the "SeventhsBass" vocabulary. Three approaches with different configurations of the system are compared with Chordino, which is probably the only both MIREX evaluated and "SeventhsBass" acceptable ACE system. Evaluation results on "The Beatles" dataset show that the best approach outperforms Chordino in the most difficult "SeventhsBass" metric in a significant way.