Note: Code for RNN model & audio synthesis is not opensourced yet.
Audio classification using deep learning implemented using TensorFlow 2.0 to detect Gunshots. 97.5% test set accuracy and 99% training set accuracy was achieved on Binary-Urban8K. This work was done during my summer internship at TUKL-NUST lab. Due to proper preprocessing & feature extraction, a simple CNN model is used to achieve promising results.
Note: scripts used to modify the data are also provided in the MISC Scripts directory.
UrbanSound8K was extended by adding 2400 gunshot files to it from AudioSet & MIVIA audio events data set.
74 more gunshots were added which were downloaded from: http://soundbible.com/tags-gun.html
Moreover, UrbanSound8K was changed for binary classification with new classes:
Finally, folds in dataset were increased from 10 to 40 to make it work on computers with less RAM memory.
dataframes_backup.h5
in Backups
sub-directory.Backups
folder named as best_weights_modified.hdf5
.