AIR Lab logo Audio Information Research Lab Electrical and Computer Engineering · University of Rochester
← All projects

Part-Invariant Model for Music Generation and Harmonization

Yujia Yan, Ethan Lustig, Joseph VanderStel, and Zhiyao Duan

Supplementary website for the paper. Paper (pdf)

Generation Examples

All examples are generated with the model trained on Bach's Chorale Dataset. The only human intervention is the instrumentation/tempo setting for the audio examples. There is no predefined key for each example.

Melody Harmonization

Four-Part Harmonization

Lemoine, Solfeo de los Solfeos 1A No. 106

Molihua

Five-Part Harmonization

Five-part harmonization (although our training set does not include any, our model can have a try).

Bassline Harmonization

In this section, we provide some generation examples on the three basslines provided in our evaluation.

Bassline 1

Example 1

Example 2

Bassline 2

Somewhat chromatic.

Example 1

Example 2

Bassline 3

Highly chromatic.

Example 1

Example 2

Longer Example

On the passacaglia theme used in Pachelbel's Canon in D

Free Generation

Example 1

Example 2

Rubrics for Objective Evaluation

TypeExplanationDeduction (pt)
1Doubled chordal 7th−1
2Doubled leading tone−1
3Parallel fifths or octaves between any voices−1
4Improperly resolved chordal 7th in any voice−1
5Improperly resolved leading tone in the soprano−1
6Non-stylistic progression (e.g., I-iii-V), or syntactic error (e.g., V-IV)−1
7Invalid chord/nontertian chord (e.g., C, C#, E, G)−1
8Melodic augmented second−0.5
9Greater than an octave between S and A, or A and T−0.5
10Voice crossing−0.5
11Voice overlap−0.5

Error Statistics

Error statistics