Weihan Xu ¹   Julian McAuley ²   Taylor Berg-Kirkpatrick ²   Shlomo Dubnov ²   Hao-Wen Dong ^2,3
1 Duke University  ² University of California San Diego  ³ University of Michigan

[Codebase]   [Paper]   [Data]  

Content

1. Selected Tag- and Text-Conditioned Generation Examples
2. Summary of MetaScore Dataset
3. Summary of Conditioned Music Generation Models
4. MetaScore Quantative Examples
5. MetaScore Annotation Examples
6. LLM Template
7. More Examples (Unselected)

Selected Tag- and Text-Conditioned Generated Music

Settings: In terms of tag-conditioned music generation, we initiate with a “start-of-song” event, then feed relevant tags into either MST-Tags-Small or MST-Tags. This is followed by a “start-of-notes” event to allow the model to begin generating notes. For text-conditioned music generation, we input the text embedding corresponding to the provided text into MST-Text and use a “start-of-song” event to prompt the model to start generating notes.

I. Multiple Control:

II. Multiple Tracks:

III. Drum Accompaniment:

IV: Free-form User-Annotated Tags:

Summary of MetaScore Dataset

Summary of Conditioned Music Generation Models

MetaScore Quantative Examples:

To evaluate the quality of our dataset, we used the rating entry (MetaScore-Raw) as an indicator. This rating, which ranges from 1 to 5 (with 5 being the highest), serves as a structured measure of perceived music quality. To demonstrate that even lower-rated entries can still be suitable for use, we randomly selected 10 samples from three categories: low ratings (below 3 or not rated), mid-range ratings (3–4), and high ratings (above 4). These examples illustrate the overall usability and diversity of the dataset across different rating levels.