Feature selection, training and validation for the ADORE sound classifier
To run the code you need:
- libXtract Python bindings, source available https://github.com/jamiebullock/LibXtract. Make sure to install Swig beforehand or Python bindings will not be built.
- sox for sound preprocessing
- Download the dataset
- Process the dataset using
prepare_audio.sh - Calculate features in a given dataset using
feature_calculation.py. Choose saving location inside the script. - Train the classifier with feature selection using
feature_selection.py. - Once you have the classifier and feature list, you can do different validation steps
- Perform cross validation on the dataset using
generate_conf_matrices_train_test.py. - Record new dataset using a robot with
play_record.py, then calculate features on that dataset (usingfeature_calculation.py) and validate the existing classifier usingclassifier_recordings.py.