| --- |
| title: MuQ-MuLan |
| emoji: 🎧 |
| colorFrom: blue |
| colorTo: green |
| sdk: gradio |
| sdk_version: 5.28.0 |
| python_version: '3.12' |
| app_file: app.py |
| pinned: false |
| license: cc-by-nc-4.0 |
| short_description: Rank music descriptions against an audio clip |
| --- |
| |
| # MuQ-MuLan Music-Text Similarity |
|
|
| A HARP-compatible deployment of MuQ-MuLan for comparing a music clip with |
| candidate text descriptions. It can support workflows such as tagging, |
| searching a sample library, or choosing the description that best matches a |
| piece of music. |
|
|
| The Space loads the official MuQ-MuLan-large checkpoint from Hugging Face. |
| Model files are downloaded at runtime and cached by the Space. |
|
|
| ## Inputs |
|
|
| - One music audio clip between 10 and 60 seconds |
| - One to eight candidate descriptions, one per line |
|
|
| Audio is converted to mono and resampled to 24 kHz. MuQ-MuLan processes |
| 10-second windows and averages their embeddings for longer clips. Candidate |
| descriptions should be written in English. |
|
|
| ## Output |
|
|
| A JSON file containing the candidate descriptions in descending similarity |
| order. Scores are cosine similarities from -1 to 1, not calibrated |
| probabilities. |
|
|
| ## Sources |
|
|
| See [SOURCES.md](SOURCES.md) for the model, source revision, license, and |
| associated paper. |
|
|
| ## License |
|
|
| The upstream source code is MIT licensed. The MuQ-MuLan model weights are |
| released under CC BY-NC 4.0, so this deployment is intended for |
| non-commercial use. |
|
|
