Using Hidden Markov Models for Musical Chord Prediction

Abstract

Music is made up of a melody and chords that accompany the melody. Finding suitable chords, can be hard and time consuming. This thesis investigates the use of Hidden Markov models (HMMs) to assist in this process. The idea is to learn the underlying Markov chain chord progressions from training data, and then adjust the chord progressions using melody input.

Four models are suggested and compared, all of them a type of HMM. Different definitions of state spaces and different orders of the models are used. Considering whole measures instead of single beats as our states, result in an improvement of the predictions. Improved predictions are also obtained by building separate models for minor key songs and major key songs. Higher order models do not improve the results. The best performing model obtains a score of 70$\%$ when using leave-one-out-cross-validation (LOOCV) for a training data set containing 64 children's songs.