SOURCES.md

Sources

MuQ and MuQ-MuLan

• Paper: https://arxiv.org/abs/2501.01108
• Source: https://github.com/tencent-ailab/MuQ
• Pinned source revision: 28847ea50cd31ac4b8b6a7dacc051ad7d1c7606a
• Source license: MIT
• Checkpoint: https://huggingface.co/OpenMuQ/MuQ-MuLan-large
• Pinned checkpoint revision: 2e01c796b71dca71b45251384c04cd7b237c9020
• Checkpoint license: CC BY-NC 4.0

MuQ-MuLan is a joint music-text embedding model trained through contrastive learning. Its audio and text embeddings are L2-normalized, so their dot product is cosine similarity. This deployment accepts English descriptions. For audio longer than 10 seconds, the official inference code embeds non-overlapping 10-second windows and averages their representations.

Encoders

• Audio encoder: https://huggingface.co/OpenMuQ/MuQ-large-msd-iter
• Pinned audio revision: 0562a57814f6f8bbd9fdea0a25921a2fce1a841a
• Audio model license: CC BY-NC 4.0
• Text encoder: https://huggingface.co/FacebookAI/xlm-roberta-base
• Pinned text revision: e73636d4f797dec63c3081bb6ed5c7b0bb3f2089
• Text model license: MIT