Explorando Modelos com Janelas Temporais para Previsão de Acordes Musicais a Partir de Melodias

The task of predicting chords in music is highly important in the careers of composers and professional musicians. Composers create new melodies that require harmony to make them, among many other things, more marketable. Professional arranger musicians frequently need to learn new songs for which they lack supporting material. This work explores chord prediction in melodies using Machine Learning. Utilizing the POP909 dataset 1, composed of 909 songs in MIDI format, Random Forest (RF) and Long Short-Term Memory (LSTM) models were trained and compared, both with and without bidirectionality, employing features such as melody, note intensity, musical key, and the delta time of each melodic interval’s execution. The RF model achieved an overall average accuracy of approximately 77%, performing well on common chords. Conversely, the LSTM without bidirectionality achieved around 61% accuracy, and with the use of the technique (BLSTM), it obtained approximately 73% accuracy. Both models demonstrated challenges in predicting rare chords, such as diminished chords. The contributions aim to advance AI-assisted musical analysis, focusing on applications for composition and live accompaniment.