Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• Existing music datasets lack clean, well-separated positive and negative examples needed for concept-based interpretability methods like TCAV.limitation 90%
• ConceptCaps is a dataset of 23,000 music-caption-audio triplets with explicit labels from a 200-attribute taxonomy.result 95%
• The dataset creation pipeline separates semantic modeling from text generation, using a VAE for attribute co-occurrence, an LLM for caption generation, and MusicGen for audio synthesis.method 90%
• This separated pipeline improves coherence and controllability over end-to-end approaches.result 85%
• The dataset is validated through audio-text alignment using CLAP.method 90%
• The dataset is validated through linguistic quality metrics including BERTScore and MAUVE.method 90%
• TCAV analysis on the dataset confirms that concept probes recover musically meaningful patterns.result 90%
• The dataset and code are publicly available.result 95%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card