mdfirman/citynet

A neural network classifier for urban soundscapes

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

doi: https://doi.org/10.1101/248708

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

Requirements

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

How to classify a new audio file with CityNet

• Run python demo.py 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:

How to classify multiple audio files

You can run CityNet on a folder of audio files with:

python multi_predict.py path/to/audio/files

This will save summaries of what is found in each wav file found to prediction_summaries.csv.

Hardware requirements

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.

citynet's Issues

Problems running on ultrasonic files

Had problems running code on files with a sampling rate of 384kHz, struggled to find where to change settings for this. It wouldn't analyse the lower frequencies. However had no issue with 48kHz files.

Theano version error (with fix)

Don't have any experience using Theano / Lasagne but I followed your dependency setup but still got an error relating to the version of Theano I was using.

I'm using WSL on Windows 10 running Anaconda with Python 2.7.13

Got it working by simply doing
pip install --upgrade https://github.com/theano/theano/archive/master.zip

Might be worth including this in the readme (if it's not just an issue specific to my setup!)