Gunshot-Detection-in-Audio

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.

Project Includes

• Project Notebook
  • Binary_Urban8K dataset visualization
  • Preprocessing and feature extraction using Librosa library
  • Pipelines for Preprocessing
  • Training using Keras, TensorFlow 2.0
  • Predictions on:
    • Test files
    • Multiple selected files
    • Real-time sound input
    • MISC Scripts Directory: it includes all the notebooks used to modify Urban8K dataset.
    • Backups Directory: it contains model weights and stored dataframe having features extracted from audio.

Dataset Details

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.

  • "UrbanSound8K.csv" was modified accordingly.

• 74 more gunshots were added which were downloaded from: http://soundbible.com/tags-gun.html

  • "UrbanSound8K-modified.csv" was created for latest version of dataset.
  • "US8K-Binary" refers to new dataset

• Moreover, UrbanSound8K was changed for binary classification with new classes:

  • no_gun_shot (8358 files, which is 3 times when compared with other class.)
  • gun_shot (2848)
  • Total files are 11206.

• Finally, folds in dataset were increased from 10 to 40 to make it work on computers with less RAM memory.

Other Details

• Currently, dataset is missing from the repository but a download link will be added soon.
• Dataframes having extracted features of dataset files are saved as dataframes_backup.h5 in Backups sub-directory.
• Model with best results is also saved in Backups folder named as best_weights_modified.hdf5.