CLSTMS: A Combination of Two LSTM Models to Generate Chords Accompaniment for Symbolic Melody

Abstract

Chords are essential in music composition, just like the frame of songs to some extent. However, choosing proper chord sequences to accompany a melody is difficult for novices, which requires the knowledge of musical structure and harmony. The paper proposes a novel method to generate chords matching a given melody using the model CLSTMS that combines two LSTM models. One model focuses on the relationship between measure notes information and corresponding chord, the other explores chord transfer rules. Experiments show that after trained on the lead sheet database, CLSTMS can learn certain principles of composing, and the performance is better than traditional HMM from the perspective of repetition rate.