Introduction
The staff notation is provided solely with melody details for ease of comprehension.
Orange: Call. Blue: Response.
Example 1: Growth ("My Heart Will Go On", James Horner)
 |
Call (melody) |
Call (melody and accompaniment) |
 |
Response (melody) |
Response (melody and accompaniment) |
Example 2: Extension ("My Heart Will Go On", James Horner)
 |
Call (melody) |
Call (melody and accompaniment) |
 |
Response (melody) |
Response (melody and accompaniment) |
Example 3: Liquidation ("My Heart Will Go On", James Horner)
 |
Call (melody) |
Call (melody and accompaniment) |
 |
Response (melody) |
Response (melody and accompaniment) |
Example 4: Conversion ("Part of Your World", Alan Menken)
 |
Call (melody) |
Call (melody and accompaniment) |
 |
Response (melody) |
Response (melody and accompaniment) |
Example 5: Condensation ("The Sound of Silence", Paul Simon)
 |
Call (melody) |
Call (melody and accompaniment) |
 |
Response (melody) |
Response (melody and accompaniment) |
Although the results suggest that human-composed music received higher ratings
than the proposed model in terms of quality, groove, and coherence, the subjects
perceived greater creativity in the music generated by CRG. It is a valuable tool
for arrangement or style transfer. Furthermore, the CRG’s ability to generate creative
responses can potentially inspire and complement human creativity during the composition
process.
Both subjective and objective experiments corroborate that our model is proficient in producing outputs with superior
musical quality.
Evidence supports that the performance of our proposed model is enhanced through the integration of a knowledge-enhanced mechanism
and multiple training tasks.
It's evident that every type of knowledge plays a part in performance improvement. Notably, knowledge candidates sharing similar composition attributes
seem to exert a more pronounced influence on enhancing model performance compared to the remaining knowledge types.
A trade-off between music diversity and music quality is observed when increasing the number of
knowledge candidates