MusCaps: Generating Captions for Music Audio

Ilaria Manco1 2, Emmanouil Benetos1, Elio Quinton2, Gyorgy Fazekas1
1 Queen Mary University of London, 2 Universal Music Group

This repository is the official implementation of "MusCaps: Generating Captions for Music Audio" (IJCNN 2021). In this work, we propose an encoder-decoder model to generate natural language descriptions of music audio. We provide code to train our model on any dataset of (audio, caption) pairs, together with code to evaluate the generated descriptions on a set of automatic metrics (BLEU, METEOR, ROUGE, CIDEr, SPICE, SPIDEr).


The code was developed in Python 3.7 on Linux CentOS 7 and training was carried out on an RTX 2080 Ti GPU. Other GPUs and platforms have not been fully tested.

Clone the repo

git clone
cd muscaps

You'll need to have the libsndfile library installed. All other requirements, including the code package, can be installed with

pip install -r requirements.txt
pip install -e .

Project structure

├─ configs                      # Config files
│   ├─ datasets
│   ├─ models  
│   └─ default.yaml              
├─ data                         # Folder to save data (input data, pretrained model weights, etc.)
│   ├─ audio_encoders   
│   ├─ datasets            
│   │   └─ dataset_name     
|   └── ...             
├─ muscaps
|   ├─ caption_evaluation_tools # Translation metrics eval on audio captioning 
│   ├─ datasets                 # Dataset classes
│   ├─ models                   # Model code
│   ├─ modules                  # Model components
│   ├─ scripts                  # Python scripts for training, evaluation etc.
│   ├─ trainers                 # Trainer classes
│   └─ utils                    # Utils
└─ save                         # Saved model checkpoints, logs, configs, predictions    
    └─ experiments
        ├── experiment_id1
        └── ...                  


The datasets used in our experiments is private and cannot be shared, but details on how to prepare an equivalent music captioning dataset are provided in the data README.

Pre-trained audio feature extractors

For the audio feature extraction component, MusCaps uses CNN-based audio tagging models like musicnn. In our experiments, we use @minzwon's implementation and pre-trained models, which you can download from the official repo. For example, to obtain the weights for the HCNN model trained on the MagnaTagATune dataset, run the following commands

mkdir data/audio_encoders
cd data/audio_encoders/
mv best_model.pth mtt_hcnn.pth


Dataset, model and training configurations are set in the respective yaml files in configs. Some of the fields can be overridden by arguments in the CLI (for more details on this, refer to the training script).

To train the model with the default configs, simply run

cd muscaps/scripts/
python <baseline/attention> --feature_extractor <musicnn/hcnn> --pretrained_model <msd/mtt>  --device_num <gpu_number>

This will generate an experiment_id and create a new folder in save/experiments where the output will be saved.

If you wish to resume training from a saved checkpoint, run

python <baseline/attention> --experiment_id <experiment_id>  --device_num <gpu_number>


To evaluate a model saved under <experiment_id> on the captioning task, run

cd muscaps/scripts/
python <experiment_id> --metrics True


      title={MusCaps: Generating Captions for Music Audio}, 
      author={Ilaria Manco and Emmanouil Benetos and Elio Quinton and Gyorgy Fazekas},