CityNet - a neural network for urban sounds

CityNet is a machine-learned system for estimating the level of biotic and anthropogenic sound at each moment in time in an audio file.

The system has been trained and validated on human-labelled audio files captured from green spaces around London.

CityNet comprises a neural network classifier, which operates on audio spectrograms to produce a measure of biotic or anthropogenic activity level.

More details of the method are available from the paper:

CityNet - Deep Learning Tools for Urban Ecoacoustic Assessment

Alison J Fairbrass, Michael Firman, Carol Williams, Gabriel J Brostow, Helena Titheridge and Kate E Jones


An overview of predictions of biotic and anthropogenic activity on recordings of London sounds can be seen at our website

Screenshot of urban sounds website


The system has been tested using:

  • Ubuntu 16.04
  • python 2.7
  • The anaconda distribution (see here for download details).

Run the following commands to get suitable versions of the required libraries:

pip install -r
pip install Lasagne==0.1
pip install --upgrade
pip install nolearn
pip install librosa
pip install easydict
pip install tqdm
pip install git+git://[email protected]

For training and testing we used a 2GB NVIDIA GPU. The computation requirements for classification are pretty low though, so a GPU should not be required.

How to classify a new audio file with CityNet

  • Run python to classify an example audio file.
  • Predictions should be saved in the folder demo.
  • Your newly-created file demo/prediction.pdf should look identical to the provided file demo/reference_prediction.pdf:

Editing should allow you to classify your own audio files.


A neural network classifier for urban soundscapes

Citynet Info

⭐ Stars 14
🔗 Homepage
🔗 Source Code
🕒 Last Update a year ago
🕒 Created 6 years ago
🐞 Open Issues 2
➗ Star-Issue Ratio 7
😎 Author mdfirman