Chord detection system using deep learning network

Author: Quang Hoang

Deep neural networks gain a huge attention from researchers recently for many successful applications using machine learning techniques, especially in natural language processing, computer vision and music data analysis problems. In this study, I focuses on using a deep learning network to analyze music data to detect and generate chord progressions from a single symbolic melody track data. In fact, the system is divided into three sub-problems including root, bass and quality detection for each note on the melody track. Each problem is solved on an equivalent neural network such as root, bass and quality network. The three networks are trained from about 775 Folk tunes (78,137 events) in the Nottingham music database [1] in ABC format and tested on about 136 songs (13,789 events) of the same database. The accuracy of the root, bass, and quality network are 59.67%, 55.26%, and 65.35% respectively. Those rates are computed by comparing exactly values of the predicted chords and the chords stored in the database. It is apparent that music is an intangible form of art, there are varying approaches to yield a chord progression for a song, it depends on the experience and feeling of composers as well as genres of the music. Therefore, this measurement method is only used for evaluating the accuracy of the training models of the networks, but measure the performance of the system. It is required to have more sophisticated methods as described in [2] to estimate the distance of chord progressions created by the system and professional musician experts. As regard to the output results of the three networks, they are combined and assigned for each event of the melody track. Based on the predicted chords and the beat strengths of events in each measure, the system synthesizes the final chord progression for the input song. For further experiments, the system will be extended to other music genres such as Classical, Blue or Jazz to generate more complicated chord progressions. This allows composers to discover new harmony arrangements from different styles of music or other composers to apply for their songs. Keyword: Harmonic Analysis, Chord Recognition, Neural Network, Deep Learning Reference: [1]. Nottingham Database. (n.d.). Retrieved November 02, 2016, from http://ifdo.ca/~seymour/nottingham/nottingham.html [2]. Arabiyat, Alaa, Al-Balqa Applied, and Mohammad F. Ababneh. "Survey-based Comparison of Chord Overlay Networks." International Journal of Computer Applications 69.12 (2013): 5-12. Print.