false_colour_index_spectrogram

Generate a false colour index spectrogram to visualise long duration soundscape recordings - based on Towsey et. al. 2014

Code tested using Python 3.12 on a 6 core Intel i7 Macbook Pro at 2.6 GHz and 14 core M3 max running MaxOS 14.4.

Install miniconda or miniforge.

Clone the repository and open the cloned folder in a terminal. The environment can be created using;

conda env create -f environment.yml

activate the newly created environment using:

conda activate false_colour_index

The code is designed to process 24 hours of audio. If this audio is placed in multiple files in the ./input_audio folder use the following command to show and save the false colour spectrograph.

python calculate_index_spectrograms.py

If you need to process many sets of 24 hours, create folders and create a batch task pointing the script at each 24 hour folder:

python calculate_index_spectrograms.py -i /path/to/your/folder1
python calculate_index_spectrograms.py -i /path/to/your/folder2
python calculate_index_spectrograms.py -i /path/to/your/folder3
python calculate_index_spectrograms.py -i /path/to/your/folder4

calculate_index_spectrograms.py contains the main functions which can be used in your own Python scripts if you import the file as a module. When using the file as a script, you can optionally change the input/output paths and you can also disable re-calculation of the statistics: run python calculate_index_spectrograms.py -h for information about the command-line options.

Usage

1. Select which indices to calculate by adding 3 function names in list indices_to_calc. Currently available are: fentropy, aci, magsum and specpow.
2. Move audio files to analyse to the ./input_audio folder or use -i command line argument to point to another location.
3. All audio files in the target folder will be combined into a single false-colour spectrogram.
4. Ensure that the naming scheme used for audio files results in consecutive files when sorted in alphabetically by name.
5. Assumes zero gap between input audio files. Any gaps between files will not be taken into account so use continuous data for each job.
6. The false-colour spectrogram will be saved using yyyy-mm-dd_HHMM_false-colour-plot.png by default. This can be changed using the --savef argument.

Command line options:

• -i, default="./input_audio", "Input path: can be a path to a single file (which will be chunked), or a folder full of wavs, or the input can be a list of wav files which you explicitly specify"
• -o, default="./default_out", "Output path: a folder (which should exist already) in which data files will be written."
• -n, default=1, choices=[0,1], "Whether to apply scaling (normalisation) of the statistics before plotting them."
• --savef, default='YYYY-MM-DD_HHMM_false_colour_plot.png', "Image file where the output figure should be saved (including extension (png, jpg, etc.). Expect issues with vector graphics")
• --hop, default=100, "How much gap between 'window', as a percentage of window size. Default of 100 %% is recommended for long >2hr. For 1hr audio you could try 25 %% for finer resolution."
• --fmin, default=0, "Lowest frequency (in Hz) to show on the plot."
• --fmax, default=44100, "Highest frequency (in Hz) to show on the plot."

plot_index_distributions.py can be used to look at the distribution of values for each of the three indices

Example output plot:

24 hours recorded from a tropical rainforest in Sabah, Borneo. Dawn and dusk choruses are clearly visible with different patterns of calling during the day and night time

Input / output data format

Input can be a single long file which will be analysed in 1-minute chunks, OR a folder containing a series of consecutive WAV files.

Long-duration recordings split into multiple WAV files of any length can be handled by this script.

24 hours of audio can be processed on a 14 core M3 with MacOS 14.4 in 31 seconds, pn a 6 core Intel i7 with MacOS 14.4 it takes 120 seconds,

Numpy ndarrays storing individual index spectrograms are stored in ./output_spectrograms/ folder.

Authors

• Sarab Sethi (Imperial College London)
• Dan Stowell (Queen Mary University of London)

Change Log

April 2024

• • Joshua Taylor (subbed in standard soundfile library and added parallel processing) (Social Environment CIC)
    • replaced chunking via modified librosa with block processing from soundfile
    • added multiprocessing ability
    • run time for 24 hours reduced from 1 hour to 32 seconds with 14-core M3 max
    • edit variable names for readability
    • runs on Python 3.12