- Python 3.6+
- LibROSA
- NumPy
- Matplotlib
- FFmpeg
- tqdm
# Setup environment
$ ./setup_env.sh -i -m
# Run unit tests
$ python3 -m unit_test.py
- The dataset was taken from the Free Music Archive (FMA): https://github.com/mdeff/fma
- The dataset consists of 8000 songs and excellent metadata that includes pre-computed features
- The FMA dataset consists of 8 genres: Electronic, Experimental, Folk, Hip-Hop, Instrumental, International, Pop, Rock
- The metadata contains various information about the songs such as artist, genre, and record date
- Pre-computed features are also part of the metadata such as MFCC, spectral contrast, and Tonnetz
- The training, validation, and test dataset sizes are 6400, 800, and 800 respectively
- The metadata for all tracks can be downloaded here: fma_metadata.zip - size is 342 MB
- A collection of Mel Spectrogram images for the dataset can be found here: fma_spectrogram.zip - size is 2 GB
- The dataset is located here: fma_small.zip - size is 7.2 GB
- Details about the FMA dataset can be found in the official paper
To access the dataset data, we can use the following APIs:
- feature_extractor class:
| Function Name | Description |
|---|---|
| get_all_song_ids | Get all song IDs in the dataset |
| get_genre | Get the genre for a song |
| get_training_dataset_song_ids | Get all song IDs in the training dataset |
| get_validation_dataset_song_ids | Get all song IDs in the validation dataset |
| get_test_dataset_song_ids | Get all song IDs in the test dataset |
| get_feature | Get feature for a song |
| get_features_as_nparray | Get features and return as a numpy array |
| get_echonest_features_as_nparray | Get Echonest features and return as a numpy array |
- audio_preprocessor class:
| Function Name | Description |
|---|---|
| save_mel_spectrogram | Save mel-scaled spectrogram image to a file |
| plot_mel_spectrogram | Plot mel-scaled spectrogram image to a file |
| get_mel_spectrogram | Get mel-scaled spectrogram as a numpy array |
| get_mel_spectrogram_with_cache | Load previously saved spectrogram if it exists. If not, it will generate spectrogram and save it as a file. |