MusicSem: 35,977 Text-Audio Music Pairings

MusicSem is a multimodal dataset containing 35,977 entries of paired text and audio. It includes a withheld test set of 480 entries for leaderboard evaluation. The dataset was curated by Rebecca Salganik, Teng Tu, Fei-Yueh Chen, Xiaohao Liu, Kaifeng Lu, Ethan Luvisia, Zhiyao Duan, Guillaume Salha-Galvan, Anson Kahng, Yunshan Ma, and Jian Kang.

Use Cases

• Training music captioning models based on paired audio and text descriptions.
• Evaluating audio-text retrieval systems based on the withheld test set.
• Developing multimodal music classification models based on the paired data structure.

Strengths

• Contains 35,977 text-audio pairs, providing a substantial corpus.
• Includes a dedicated test set of 480 entries for benchmarking.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the main dataset is known, but other metadata like file formats and size is unknown.

Provenance

Source

AMSRNA via Hugging Face.

Freshness

Last updated 2025-05-15 23:40:55.