Skip to main content

Audio Cover Song Suite (acoss): A feature extraction and benchmarking suite for cover song identification tasks

Project description

acoss: Audio Cover Song Suite

PyPI pyversions Build Status License: AGPL v3

acoss: Audio Cover Song Suite is a feature extraction and benchmarking frameworks for the cover song identification (CSI) tasks. This tool has been developed along with the new DA-TACOS dataset.

acoss includes a standard feature extraction framework with state-of-art audio features for CSI task and open source implementations of seven state-of-the-art CSI algorithms to facilitate the future work in this line of research. Using this framework, researchers can easily compare existing algorithms on different datasets, and we encourage all CSI researchers to incorporate their algorithms into this framework which can easily done following the usage examples.

Please site our paper if you use this tool in your resarch.

Furkan Yesiler, Chris Tralie, Albin Correya, Diego F. Silva, Philip Tovstogan, Emilia Gómez, and Xavier Serra. Da-TACOS: A Dataset for Cover Song Identification and Understanding. In 20th International Society for Music Information Retrieval Conference (ISMIR 2019), Delft, The Netherlands, 2019.

Benchmarking results on Da-Tacos dataset can be found in the paper.

Setup & Installation

We recommend you to install the package inside a python virtualenv.

Install using pip

pip install acoss


Install from source (recommended)

  • Clone or download the repo.
  • Install acoss package by using the following command inside the directory.
python3 install

Additional dependencies

acoss requires a local installation of madmom library for computing some audio features and essentia library for similarity algorithms.

For linux-based distro users,

pip install "acoss[extra-deps]"

or if you are a Mac OSX user,

you can install madmom from pip

pip install madmom

you can install essentia from homebrew

brew tap MTG/essentia
brew install essentia --HEAD

Documentation and examples

acoss mainly provides the following python sub-modules-

  • acoss.algorithms: Sub-module with various cover identification algorithms, utilities for similarity comparison and an template for adding new algorithms.

  • acoss.coverid: Interface to benchmark a specific cover identification algorithms.

  • acoss.features: Sub-module with implementation of various audio features.

  • acoss.extractors : Interface to do efficient batch audio feature extraction for an audio dataset.

  • acoss.utils : Utility functions used in acoss package.


Cover Song Identification algorithms in acoss
Serra09 Paper
LateFusionChen Paper
EarlyFusionTraile Paper
SiMPle Paper
FTM2D Paper
MOVE adding soon ...


	"chroma_cens": numpy.ndarray,
	"crema": numpy.ndarray,
	"hpcp": numpy.ndarray,
	"key_extractor": {
		"key": numpy.str_,
		"scale": numpy.str_,_
		"strength": numpy.float64
	"madmom_features": {
		"novfn": numpy.ndarray, 
		"onsets": numpy.ndarray,
		"snovfn": numpy.ndarray,
		"tempos": numpy.ndarray
	"mfcc_htk": numpy.ndarray,
	"label": numpy.str_,
	"track_id": numpy.str_

Dataset structure required for acoss

Audio data


Feature file

import deepdish as dd

feature = dd.load("feature_file.h5")

An example feature file will be in the following structure.

   'feature_1': [],
   'feature_2': [],
   'feature_3': {'type_1': [], 'type_2': [], ...},

CSV annotation file for a dataset

work_id track_id
W_163930 P_546633
... ...

acoss benchmark methods expect the dataset annotation csv file in the above given format.

There are also some utility functions in acoss which generates the csv annotation file automatically for da-tacos from it's subset metadata file and for covers80 dataset from it's audio data directory.

from acoss.utils import da_tacos_metadata_to_acoss_csv

from acoss.utils import generate_covers80_acoss_csv

For quick prototyping, let's use the tiny covers80, dataset.

Audio feature extraction

from acoss.utils import COVERS_80_CSV
from acoss.extractors import batch_feature_extractor
from acoss.extractors import PROFILE


{ 'sample_rate': 44100, 'input_audio_format': '.mp3', 'downsample_audio': False, 'downsample_factor': 2, 'endtime': None, 'features': ['hpcp', 'key_extractor', 'madmom_features', 'mfcc_htk'] }


# Let's define a custom acoss extractor profile
extractor_profile = {
           'sample_rate': 32000,
           'input_audio_format': '.mp3',
           'downsample_audio': True,
           'downsample_factor': 2,
           'endtime': None,
           'features': ['hpcp']

# path to audio data
audio_dir = "../coversongs/covers32k/"
# path where you wanna store your data
feature_dir = "features/"

# Run the batch feature extractor with 4 parallel workers


from acoss.coverid import benchmark, algorithm_names
from acoss.utils import COVERS_80_CSV

# path to where your audio feature h5 files are located
feature_dir = "features/"

# list all the available algorithms in acoss 

# we can easily benchmark any of the given cover id algorithm
# on the given dataset using the following function
# result of the evaluation will be stored in a csv file 
# in the current working directory.

How to add my algorithm in acoss?

Defining my algorithm

from acoss.algorithms.algorithm_template import CoverSimilarity

class MyCoverIDAlgorithm(CoverSimilarity):
    def __init__(self, 

    def load_features(self, i):
        """Define how to want to load the features"""
        feats = CoverAlgorithm.load_features(self, i)
        # add your modification to feature arrays

    def similarity(self, idxs):
        """Define how you want to compute the cover song similarity"""

Running benchmark on my algorithm

# create an instance your algorithm
my_awesome_algorithm = MyCoverIDAlgorithm(dataset_csv, datapath)

# run pairwise comparison

# Compute standard evaluation metrics
for similarity_type in my_awesome_algorithm.Ds.keys():


How to contribute?

  • Fork the repo!
  • Create your feature branch: git checkout -b my-new-feature
  • Add your new audio feature or cover identification algorithm to acoss.
  • Commit your changes: git commit -am 'Add some feature'
  • Push to the branch: git push origin my-new-feature
  • Submit a pull request


This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 765068 (MIP-Frontiers).

This work has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No 770376 (Trompa).

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

acoss-0.0.3.tar.gz (123.2 kB view hashes)

Uploaded source

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page