MusicBench: 53,168 Audio-Text Pairs for Music Generation

MusicBench is a music audio-text pair dataset designed for text-to-music generation. It expands the MusicCaps dataset from 5,521 to 52,768 training and 400 test samples. The dataset was released by amaai-lab in March 2025 alongside the Mustango model.

Use Cases

• Train text-to-music generation models based on the audio-text pairs.
• Fine-tune music language models based on the expanded sample count.
• Analyze relationships between musical features and descriptive text based on the included chords, beats, tempo, and key data.

Strengths

• Substantially expands the MusicCaps dataset to 52,768 training samples.
• Includes extracted music features such as chords, beats, tempo, and key.
• Provides a dedicated test set of 400 samples for evaluation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full dataset is not explicitly stated, only training/test splits.
• Data may reflect bias inherent to the source MusicCaps dataset.

Provenance

Source

amaai-lab on Hugging Face, based on the MusicCaps dataset.

Collection Method

Expansion of an existing dataset with additional feature extraction.

Freshness

Last updated 2025-03-20 09:34:35