Spplication to convert bird sound analysis made with BirdNET into a more usable format:

• Excel file with sheets for
  • Predictions with filename, and position in h:mm:ss format, to make finding the sound eaier
  • Number of predictions per species above given threshold, to allow seeing what are the common vs. less common species
• HTML report page with five audio clips of each species. This makes it easy to check if predictions for each species is reliable or not.

Setup

Initial setup

git clone https://github.com/mikkohei13/baim.git
apt install python3.10-venv
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
(install Tkinter)

Start virtual environment

source .venv/bin/activate
deactivate

Using from command line

• Analyze audio files with BirdNET
• Have birdNET result files in a root directory and audio files in Data directory
• Set path to directory to handle_files.py
• Run ´python3 handle_files.py´

Compiling the app on Windows

• Setup Python on Windows

• Set Python to Powershell profile

• Copy files to Windows (no subdirectories): ´cp ./baim/* /mnt/c/Users/mikko/Documents/compile/´

• With Powershell:

  • Install requirements: ´pip install -r requirements.txt´
  • Compile the app:

  pyinstaller --windowed --onefile app.py --add-data "C:/Users/mikko/AppData/Local/Programs/Python/Python310/Lib/site-packages/librosa/util/example_data/registry.txt;librosa/util/example_data" --add-data "C:/Users/mikko/AppData/Local/Programs/Python/Python310/Lib/site-packages/librosa/util/example_data/index.json;librosa/util/example_data" --add-data "baim-icon.png;." --distpath "C:/Users/mikko/Desktop/" --hidden-import "sklearn.metrics._pairwise_distances_reduction._datasets_pair" --hidden-import "sklearn.metrics._pairwise_distances_reduction._middle_term_computer"

How it works

Creating audio snippets & spectrograms of species

• Have list of non-finnish species
• Have a prediction dataframe
• Filter confidence >= 0.9
• Set empty dict for predictions-to-check
• Set empty dict with scientific name as key, count as value
• Loop rows
  • If species has < 5 occurrences in names dict
    • If species is not non-finnish
      • Add to preditions dict
      • Add tp names dict
• Filter confidence <= 0.7
  • Do same loop as above
• Now we have dict of max 5 snippets per species
• Sort by scientific name asc, confidence asc (or time asc?)
• Loop the dict
  • Get start time & end time
  • Set cut_start == start - 5, cut_end == end + 5.
  • If cut_start < 0, set cut_start == 0
  • If cut_end > audio file len, set cut_end == len
  • Cut audio snippet from cut_start to cut_end
  • Make spectrogram, with width based on length in seconds
  • Save audio and spectro in report file
  • Generate html for the snippet, using
    • audio
    • spectro
    • species, confidence
    • filename
    • times (also in ISO format that Vihko uses)
• Save html report

Expectations

• BirdNET files are in the csv format provided by birdnet
• All audio files have same file extension. This can be wav, mp3 or flac
• Audio filenames are in format created by Audiomoth (3 versions) or Wildlife Acoustics SM4

Todo

DONE: Randomixe dataframe order before picking segments? To avoid having 5+ segments of the same local bird.

To get most probable identifications, sort by confidence desc. To get a sample of all identifications above the threshold, shuffle, to get good idea of the variety.

Exceliin sarakkeiksi myös: Laji - Määritys Määrä - Havainto Pesimävarmuusindeksi - Havainto Lisätiedot - Havainto Kokoelma/Avainsanat - Havainto

Ideas

Filter low-frequency noise, unless predicted species has low frequency (botste, owls...)