This directory contains all the scripts and data files to replicate our results. Some scripts may require you to register at Freesound.org for a free API key.
Most people are probably interested in our models. The folder models contains
a file called tag-lsa.py that contains all the code to build a tag-based distributional model. This model performs pretty well on the MEN-dataset, as you can see from the results using test_suite.py. We made this file as general as possible, so you can evaluate your own models using the same code. The evaluate_model function takes a matrix, list of row labels, and the name of the test set as its argument. It returns a dictionary with the results. The evaluate_word2vec function takes a word2vec model (created using Gensim), and the name of the test set as its argument. Its output is the same. For the sound models, there are two scripts inside the models folder: sound_database.py preprocesses the data and splits it into a training set and validation set. BoAW.py contains the code to create the BoAW-models.
Please unzip all_freesound_data.csv.zip if you are planning to use scripts making use of this resource.
###Important
- Credit the creators of the sounds if you plan on using the sounds in any of your projects. Be aware that some sounds may have a more restrictive license than others.
- Please consider donating to the Freesound project.
To run all of our experiments, you will need the following packages:
| Package | What does it do? |
|---|---|
| Gensim | Topic modeling |
| Networkx | Network interface |
| numpy | Math |
| python-louvain | Network analysis |
| scikit-learn | Machine learning |
| unicodecsv | Unicode I/O |
| tabulate | LaTeX results table |
| pydub | Sound processing toolbox |
| features | Sound processing toolbox |