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 londonsounds.org.
The system has been tested using the dependencies in environment.yml. Our code works with python 3.
You can create an environment with all the dependencies installed using:
conda env create -f environment.yml -n citynet
conda activate citynet- Run
python demo.pyto classify an example audio file. - Predictions should be saved in the folder
demo. - Your newly-created file
demo/prediction.pdfshould look identical to the provided filedemo/reference_prediction.pdf:
You can run CityNet on a folder of audio files with:
python multi_predict.py path/to/audio/filesThis will save summaries of what is found in each wav file found to prediction_summaries.csv.
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.