Identify bird calls using AI, and monitor call frequency.

Skraak is intended to be simple to use for simple people like me.

This package serves

Most of the data has been recorded using Open Acoustics AudioMoth's or μMoth's at 16000 Hz. DOC recorders at 8000hz work fine.

It is a good idea to use an Nvidia GPU. Everything should work fine on CPU, just slow.

AMD and Mac Silicone GPU's are not supported but should be easy for you to get working with julia AMD or Metal packages.

If you are doing serious work, start the julia repl with: julia -t n where n is up to 1/2 the number of cores you have. I do 4, this is enough to keep up with a gamer style GPU.

You can use Skraak too.

[Install Julia](, Julia-1.10 or newer
[git clone the Skraak project](, if you dont have git or the git cli, you can download a zip file by clicking the <code> button.

cd to your Skraak folder
start the julia repl with $julia
(You will want to install Revise and OhMyREPL, just do 'using Revise, OhMyREPL' in the Julia repl, add 'using Revise, OhMyREPL' to ~/.julia/config/startup.jl)

type: ] (to enter Pkg mode)
type: activate .
type: instantiate
backspace to exit Pkg mode
exit repl with ctrl-D.


start the julia repl with $julia
type: ] (to enter Pkg mode)
type: dev path/to/Skraak (to make it a local package)
backspace to exit Pkg mode
type: using Skraak, Glob (glob is only here to help you refine your glob patterns)
When finished working you can if you like do 'free Skraak' in Pkg mode (accessed with ']')
  1. Take some WAV's organised into a file structure LOCATION/TRIP_DATE/WAV_FILES
  2. and labels saved in a csv in the form:
    • file(String),start_time,end_time,label(Integer) (where start_time and end_time are in seconds from the start of the wav file)
    • at least 2 label classes are required, for example Kiwi, Not
  3. Generate a primary dataset of spectrogram images with the following file structure:
    • DATASET/AUDIO_FILE*/LABEL*/PNG's (png files must be 224X224 px square RGB).
    • This structure is required, when training, the parent folder of a file is the label.
    • This function creates a folder for each file, creates subfolders for each label, then saves png files in the appropriate label sub folder.
    • Space is needed. It uses the whole audio file. (I aim for 96% Not, 4% Kiwi)
    • and saves a flac copy for reference

I use labels, [K, N] in words [Kiwi, Not]. Anything will work, the unique text labels are sorted alphabetically and mapped to integer labels in the training process. More than 2 label classes is fine, but keep it simple until you have a lot of data. It is better not to have everything in big folders, 100_000 files in a folder on a Fat32 removable drive will rapidly grind to a stand still. You could have many thousands of K and N folders, for example, the model does not care. Native file systems on mac/linux will work ok. I use ext4 (linux) file systems on exteranl SSD's for both linux and mac.

  1. Train a Resnet18 model, either pretrained on Imagenet, or preferably the pretrained Skraak Kiwi model, which is currently trained on 7_700_000 images. Skraak trains on 5 second clips, converted to 224x224 pixel RGB spectrogram images.
using Skraak
glob_pattern_1 = "Clips*/[D,F,M,N]/*.png" #for example. Note: requires png's as input.
glob_pattern_2 = "Dataset*/[K, N]/*.png"

# Train a model named Test1 for 2 epochs on png files found by glob_pattern, 
# start with a pretrained model.
train("Test1", 2, glob_pattern_1, true) 

# Train a model named Test2 for 2 epochs on png files found by glob_pattern, 
# train using model found at "path/to/model.jld2"
train("Test2", 2, glob_pattern_2, "path/to/model.jld2") 

# Note: Your unique text labels are sorted alphabetically, and converted to  
# integers, [1,2,3...] to be consumed by the flux model
# A text file will be saved beside the model.jld2, with the label to 
# integer mapping.
  1. Run inference on raw data using a trained model Skraak will try to find png images first, in the folders covered by the glob pattern. If there are no png's found it will predict on wav or flac files, using 5 second audio clips, converted to 224x224 pixel RGB spectrogram images, with a 2.5 second hop.

You are responsible for providing an appropriate model. I use a binary Kiwi/Not model for finding calls in audio data, and a Duet/Female/Male/Not model on png clips made from calls detected by the binary model. Find some models to start with in the Models folder

using Skraak
glob_pattern = "*/*/" #Note: requires folders as input. Folders contain flac, wav or png files.

# Predict label classes of png, wav or flac files found in folders specified by 
# glob_pattern using model.jld2. A preds.csv file is saved in current directory
predict(glob_pattern, "path/to/model.jld2")
  1. Generate audio clips and spectrogram images of all calls found.
# Make clips from a preds.csv file of the form:
# file(String),start_time,end_time,label(Int)
# 1 is the label, it can be any int present in the label field of preds.csv
# It saves clips in a folder 'Clips_2023-11-09'
make_clips("preds.csv", 1)
  1. Sort calls into subclasses (say: Duet, Female, Male, Nothing) manually, or using a model combined with human supervision. TODO

  2. Store data from calls and file metadata in a DuckDB database for statistical analysis using SQL, DataFrames, Plots.

I will not document this until the DuckDB storage api has stabilised. 
For now always store a csv backup using "EXPORT DATABASE 'Backup_2023-10-10';" in the duckdb cli.
I highly recommend storing data in a duckdb database.
Querying a duckdb database with SQL is faster than even julia DataFrames, both leave Pandas in the dust.
  1. Repeat, iterating on your models as you accumulate more data. It's hard until it gets easy.

Managing datasets is like gardening, it takes some weeding and a lot of compost (aka data) to get a good model growing.

Julia is great for machine learning because it is realtively simple to get a GPU working. It does have disadvantages at GPT-4 scale, but for this kind of work it is excelent. Julia shines with any scientific computing task.