Environmental Sound Classification

Classification of urban sounds using deep learning.

Usage

Requires Python3+ and Spark 2.0+.

git clone [email protected]:bhavika/UrbanSound.git

cd UrbanSound

virtualenv venv

source venv/bin/activate

pip install -r requirements.txt

chmod +x run.sh

./run.sh

Models

There are 2 models built using Keras and Tensorflow - they are in src/cnn.py and src/sbcnn.py. The CNN is a simple 2 layer neural network, whereas sbcnn.py contains an implementation of the SBCNN model from this paper.

You can run train the CNN and predict on 3 folds of the UrbanSound8K dataset using python3 src/cnn.py

Similarly, to run the SBCNN - python3 src/sbcnn.py.

We've also implemented a data-distributed model training setup using Elephas. This is shown in src/dist_sbcnn.py. We use 2 workers on an m4.2xlarge instance to achieve data-distributed training. If you have Spark set up, you can test this locally using - spark-submit src/dist_sbcnn.py.

The runtime for each of these models can be anywhere from a few minutes (15 minutes for dist_sbcnn.py with 200 epochs to a few hours (sbcnn.py with all configurations).

Troubleshooting

1. On Debian systems, you might run into issues with tensorboard if you don't have tkinter installed. This can be resolved by installing python3-tk.

sudo apt-get install python3-tk

2. Librosa requires an audio backend for processing WAV files (in UrbanSound8k/audio). If you see errors that indicate the absence of this backend, you might be missing libav-tools on Debian. Install them with -

sudo apt-get install libav-tools