1. python -m venv venv
2. source venv/bin/activate
3. pip install -r requirements.txt

1. Download FMA small dataset (8k tracks, 30s each) from https://os.unil.cloud.switch.ch/fma/fma_small.zip and unzip in ./data
2. python build_dataset.py (this will take a lot of time, but can be stopped and resumed at any moment)
3. python train.py (see python train.py --help for options, can be stopped and resumed)
4. (optional - evaluation) python test.py --small_encoder <path to small encoder state> --large_encoder <path to large encoder state>

python indexer.py --small_encoder <path to small encoder state> --large_encoder <path to large encoder state> --database <where to save track database> --index <where to save lookup index> --data <directory containing your mp3 files> ( this will take a lot of time)

Use recognition.py (see python recogonition.py --help for options).

This tool works either in an interactive mode (audio sample paths, offsets and lengths provided via stdin) or can be pointed to a directory containing mp3 files (random samples will be taken). Output the name of the closest matching track in the database (and similarity score - the lower the better).

• Recognize audio samples directly from pulseaudio.
• Split recognition.py into a client and a server (data loads slowly).