Automatic melody harmonization with triad chords: A comparative study

Melody Harmonization Using Orderless Nade, Chord Balancing, and Blocked Gibbs Sampling

“…where w s is the weight of chord s [22], x chordt,s is the s-th element of the one-hot vector x chordt , C is the context given, and θ denotes the model parameters. Chord weighting can greatly solve the problem of excessive use of common chords in Lim's model [9] like Yeh's chord function [12]. However, our model can further manage to assign appropriate altered chords when the melody requires a secondary or borrowed chord to elaborate the whole piece, but Yeh's model [12] cannot.…”

“…Chord weighting can greatly solve the problem of excessive use of common chords in Lim's model [9] like Yeh's chord function [12]. However, our model can further manage to assign appropriate altered chords when the melody requires a secondary or borrowed chord to elaborate the whole piece, but Yeh's model [12] cannot. The weight matrix is derived by…”

“…For objective evaluation, we use six different objective metrics introduced in [12]. The tonal distance proposed in [24] for measuring the distance between two chords.…”

“…The first three metrics measure the quality of the chord progression, and the others measure the coherence between the melody and chords. For details about the metrics, see [12]. For subjective evaluation, we randomly selected 50 melodies from our test set.…”

“…Based on the same neural structure, Yeh et al extended the number of chords to 48 by considering major, minor, augment and diminish chords. In addition, they integrated information of chord functions [11] into the loss function to help chord label prediction, and predicted a chord every half bar [12]. Experiments showed that their model, called MTHarmonizer, could generate more creative and elaborated chord progressions than the previous model.…”