Silent Cities AudioTagging

Applying a pretrained DL model to annotate soundscapes. This was developed for analyzing the data collected in the Silent Cities project.

Used model:

• Audio Tagging on Audioset using ResNet22 from PANNs

Requirements

• pytorch 1.4.0
• librosa
• tqdm
• pandas
• numpy
• matplotlib
• plotly

Usage

python tag_silentcities.py [-h] [--length LENGTH] 
                   [--folder FOLDER] [--file FILE] [--verbose]
                   [--overwrite] [--out OUT]

Silent City Audio Tagging with pretrained LeeNet11 on Audioset

optional arguments:
-h, --help       show this help message and exit
--length LENGTH  Segment length
--folder FOLDER  Path to folder with wavefiles, will walk through subfolders
--file FILE      Path to file to process
--verbose        Verbose (default False = nothing printed)
--overwrite      Overwrite files (default False)
--out OUT        Output file (pandas pickle), default is output.xz

This will save a pandas dataframe as an output file.

A heatmap can be generated using the function in analysis.py and postprocessing.py, to generate a heatmap such as this one :

Use the make_interactive_pdf function to generate an estimate of probability densities at various scales, such as this one :

Credits

Nicolas Farrugia, Nicolas Pajusco, IMT Atlantique, 2020.

Code for Audioset Tagging CNN from Qiu Qiang Kong