Neural audio fingerprinting model for Music Informatics course at QMUL. Inspired by the SimCLR approach, similar to: Chang, Sungkyun, et al. "Neural audio fingerprint for high-specific audio retrieval based on contrastive learning." ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021. but with a more lightweight encoder, simpler nearest neighbor search, and implemented in PyTorch.
First, you need to install the requirements in a python environment:
pip install -r requirements.txt
Otherwise, you could try your luck with an existing environment that has torch, torchaudio, librosa, numpy, and scikit-learn.
Then, you can run the fingerprint database builder in your python file:
from main import FingerprintBuilder
FingerprintBuilder(</path/to/database/>, </path/to/fingerprints/>)and the identification process using:
from main import audioIdentification
audioIdentification(</path/to/queryset/>, </path/to/fingerprints/>, </path/to/output.txt>)You additionally need to install weights and biases for tracking in the mentioned environment.
pip install wandb
You don't have to create a wandb account, you can choose to save logs locally only when you launch training.
You can launch training using
python train.py
You would need to have your training dataset in ./data/database_recordings, and the background mix files in ./bg, which are already provided in this repository.