Large Scale Cover Detection in Digital Music Libraries using Metadata, Lyrics and Audio Features

Source code and supplementary materials for the paper "Correya, Albin, Romain Hennequin, and Mickaël Arcos. "Large-Scale Cover Song Detection in Digital Music Libraries Using Metadata, Lyrics and Audio Features." arXiv preprint arXiv:1808.10351 (2018)".

This repo contains scripts to run text-based experiments for cover song detection task on the MillionSongDataset (MSD) which is imported into an Elasticsearch (ES) index as described in the above mentioned paper.

Requirements

Install python dependencies from the requirements.txt file

$ pip install -r requirements.txt

Setup

• Use ElasticMSD scripts to setup your local Elasticsearch index of MSD.
• Fill your ES db credentials (host, port and index) as a environment variable in your local system. Check templates.py file.

Datasets

The following datasets have corresponding mapping with MSD tracks. These data are ingested to the ES index in an update operation

• Second Hand Songs (SHS) dataset. Check the ./data folder
• For lyrics we used the musiXmatch (MXM) dataset

Usage

Modular mode

In this section, you can have a glimpse on how to use these classes and various methods for doing experiments

#import modules
from es_search import SearchModule
from experiments import Experiments
import templates as presets

# Initiaite es search class
es = SearchModule(presets.uri_config)

# search method by msd_track title in view mode
results = es.search_by_exact_title('Listen To My Babe', 'TRPIIKF128F1459A09', mode='view')

#You can also use the experiment class to automate particular experiments for a method
#Initiate experiment class with the instance of SearchModule and path to the dataset as arguments
exp = Experiments(es, './data/test_shs.csv')

#run the song title match experiment with top 100 results
results = exp.run_song_title_match_task(size=100)

#compute evaluation metrics for the task
mean_avg_precison = exp.mean_average_precision(results)

#reset the preset if you want to do another experiment on the same same SearchModule instance.
exp.reset_preset()

results = exp.run_mxm_lyrics_search_task(size=1000)

mean_avg_precison = exp.mean_average_precision(results)

Evaluation tasks

Some examples for using functions in evaluations.py script to reproduce the results mentioned in the paper

from evaluations import *

#Evaluation task on SHS train set against the whole MSD (1 x 999,999 songs)
shs_train_set_evals(size=100, method="msd_title", mode="msd", with_duplicates=True)

#You can specify various prune sizes and methods as parameters
shs_train_set_evals(size=1000, method="mxm_lyrics", mode="msd", with_duplicates=False)

#You can run the same experiment only on the SHS train set against itself by specifying "mode" param as "shs" (1 x 12,960)
shs_train_set_evals(size=100, method="msd_title", mode="shs", with_duplicates=True)

#In same way you can do the evaluation experiments on SHS test sets
shs_test_set_evals(size=100, method="title_mxm_lyrics", with_duplicates=True)

If you don't want to care about how the module works and you only need results various experiments, then this is for you. It's a wrapper around the modules to run automated experiments and save the results to a .log file or a json_template. The experiments are multi-threaded and able to run from terminal using command-line arguments.

$ python evaluations.py -m test -t -1 -e msd -d 0 -s 100

    -m : (type: string) Choose between "train" or "test" modes
    -t : (type: int) No of threads
    -e : (type: int) Choose between "msd"
    -d : (type: boolean) include duplicates
    -s : (type: int) Required pruning size for the experiments

Cite

If you use these work, please cite our paper.

Correya, Albin, Romain Hennequin, and Mickaël Arcos. "Large-Scale Cover Song Detection in Digital Music Libraries Using Metadata, Lyrics and Audio Features." arXiv preprint arXiv:1808.10351 (2018).

Bibtex format

@article{correya2018large,
  title={Large-Scale Cover Song Detection in Digital Music Libraries Using Metadata, Lyrics and Audio Features},
  author={Correya, Albin and Hennequin, Romain and Arcos, Micka{\"e}l},
  journal={arXiv preprint arXiv:1808.10351},
  year={2018}
}