Official repository of PeakNetFP: Peak-based Neural Audio Fingerprinting Robust to Time Stretching by Guillem Cortès-Sebastià, Benjamin Martin, Emilio Molina, Xavier Serra, and Romain Hennequin. To be presented at ISMIR 2025. Preprint is available in arXiv.
PeakNetFP runs in a docker container. The same container can be used to run PeakNetFP and NeuralFP.
-
Clone the repository.
-
Build docker image. Modify the
docker-compose.ymlif we want to change the image or containers names, for instance.docker-compose build
-
Build and run the container.
docker-compose up -d
-
Open an interactive shell
docker exec -it peaknetfp /bin/bash
To reproduce the experiments in this publication two data sets are needed:
We use the dataset from Neural-Audio-FP (https://github.com/mimbres/neural-audio-fp?tab=readme-ov-file#dataset) for training and validation. The scalability of the approach is tested by adding the tracks of test-dummy-db-100k-full to the reference DB. Hence the importance to download the full neural-audio-fp dataset, available in IEEEDataPort (https://ieee-dataport.org/open-access/neural-audio-fingerprint-dataset). We train with train-10k-30s subset while val-query-db-500-30s is used for validation.
The test dataset is created in this publication from the test set of neural-audio-fp test set and it is available in 
test-query-db-500-30s stretched by the following factors: 0.5, 0.6, 0.7, 0.8, 0.9, 0.95, 0.975, 1 (no stretch), 1.05, 1.1, 1.2, 1.4, 1.6, 1.8, and 2. using the tempo functionality of SOX (https://sourceforge.net/projects/sox/). The downloaded query_stretch folder must be placed inside the test-query-db-500-30s folder of the neural-audio-fp dataset. So the tree structure of the dataset directory should look like this
neural-audio-fp-dataset
├── aug
├── extras
├── LICENSE
├── music
│ ├── LICENSE-fma
│ ├── others
│ ├── test-dummy-db-100k-full
│ ├── test-query-db-500-30s
│ │ ├── db
│ │ ├── query
│ │ ├── query_stretch --> PeakNetFP stretched test set
│ │ ├── test_ids_icassp2021.npy
│ ├── train-10k-30s
│ └── val-query-db-500-30s
└── README.md
Model checkpoints are available in 
logs/checkpoint.
- Check batch-size that fits on your device first.
- Check the config file and adjust the parameters accordingly.
python run.py train -c config/<config_file.yml>example with stretching 0.9
python run.py generate <exp_name> <ckpt_index> \
-c config/<config_file.yaml> \
-o logs/emb/<exp_name>/tempo-0_900/<ckpt_index> \
--query_dir /datasets/neural-audio-fp-dataset/music/test-query-db-500-30s/query_stretch/tempo-0_900It will generate the following files:
.
└──logs
└── emb
└── CHECKPOINT_NAME
└── CHECKPOINT_INDEX
├── db.mm
├── db_shape.npy
├── dummy_db.mm
├── dummy_db_shape.npy
├── query.mm
├── query_shape.npy
├── query_segments.csv
├── db_segments.csv
└── dummy_db_segments.csv By default config, generate will generate embeddings (or fingerprints)
from 'dummy_db', test_query and test_db. The generated embeddings will
be located in logs/emb/<exp_name>/<ckpt_index>/**.mm and
**.npy.
dummy_dbis generated from the test set specified in the config file.- In the
DATASELsection of config, you can select options for a pair ofdbandquerygeneration. The default isunseen_icassp, which uses a pre-defined test set. - It is possilbe to generate only the
dbandquerypairs by--skip_dummyoption. This is a frequently used option to avoid overwriting the most time-consumingdummy_dbfingerprints in every experiment. - It is also possilbe to generate embeddings (or fingreprints) from your
custom source using
-sor--sourceargument during generation.
example with stretching 0.9
- PeakNetFP performs song-level search, but we keep the segment level search code from NeuralFP for backwards compatibility.
python run.py evaluate <exp_name> <ckpt_index> \
-c config/<config_file.yaml> \
--emb_dir logs/emb/neuralfp/tempo-0_900/<ckpt_index> \
--stretch_factor 0.9Run a dedicated docker container to check tensorboard:
docker run --network=host -ti --rm --name tensorboard -v <vol_path>:<vol_path> tensorflow/tensorflowthen run:
tensorboard --logdir <vol_path>/neuralfp/logs/fit --port <port> --bind_allMain developments with respect to the original NeuralFP repository:
- Fix some parts of the code such as dataset generation loader.
- Correct the data normalization (now each melspec is normalized, before it was normalize at batch level).
- Upgrade packages to newer versions. This was motivated by the fact that the original code with tensorflow==2.4.1 cannot be run in newer gpu cards, so we upgraded to the last version 2.15.
- ADD peak spectrogram, gaussian spectrogram transformations to input data.
- ADD stretching augmentation
- Include PointNet++ code to build PeakNetFP
This research is part of resCUE – Smart system for automatic usage reporting of musical works in audiovisual productions (SAV-20221147) funded by CDTI and the European Union - Next Generation EU, and supported by the Spanish Ministerio de Ciencia, Innovación y Universidades and the Ministerio para la Transformación Digital y de la Función Pública. Furthermore, it has received support from the Industrial Doctorates plan of the Secretaria d’Universitats i Recerca, Departament d’Empresa i Coneixement de la Generalitat de Catalunya, grant agreement No. DI46-2020.
To be updated once the proceedings are published
G. Cortès-Sebastià, B. Martin, E. Molina, X. Serra, R. Hennequin, “PeakNetFP: Peak-based Neural Audio Fingerprinting Robust to Extreme Time Stretching”, in Proc. of the 26th Int. Society for Music Information Retrieval Conf., Daejeon, South Korea, 2025.
@article{cortes2025peaknetfp,
title={PeakNetFP: Peak-based Neural Audio Fingerprinting Robust to Extreme Time Stretching},
author={Cort{\`e}s-Sebasti{\`a}, Guillem and Martin, Benjamin and Molina, Emilio and Serra, Xavier and Hennequin, Romain},
journal={arXiv preprint arXiv:2506.21086},
year={2025},
note={Accepted at ISMIR 2025}
}