Skip to content

Latest commit

 

History

History
126 lines (88 loc) · 9.04 KB

create_dataset.md

File metadata and controls

126 lines (88 loc) · 9.04 KB
Raw

Instructions for setting up datasets

In out paper we analyse four datasets:

  • FSL4: This dataset contains user-crontributed loops uploaded to Freesound. It has been built in-house by searching Freesound for sounds with the query terms loop and bpm, and then automatically parsing the returned sound filenames, tags and textual descriptions to identify tempo annotations made by users. For example, a sound containing the tag 120bpm is considered to have a ground truth of 120 BPM.
  • APPL: This dataset is composed of the audio loops bundled in Apple's Logic Pro music production software. We parsed the metadata embedded in the audio files using source code available in a public repository, and extracted in this way tempo annotations for all the loops.
  • MIXL: This dataset contains all the loops bundled with Acoustica's Mixcraft 7 music production software. Tempo annotations are provided in its loop browser and can be easily exported into a machine-readable format.
  • LOOP: This dataset is composed of loops downloaded from Looperman, an online loop sharing community. It was previously used for research purposes in [1]. Tempo annotations are available as metadata provided by the site.

[1] Gerard Roma. Algorithms and representations for supporting online music creation with large-scale audio databases. PhD thesis, Universitat Pompeu Fabra, 2015.

Due to license restrictions, we can only directly provide the contents of FSL4 dataset. We provide two ways of getting the dataset, either by downloading a single zip file or running a script that gets the contents from Freesound. Instructions for both options can be found below.

Two of the other datasets, APPL and MIXL, are in general easy to reproduce if owning the corresponding music production software. You'll also find instructions below on how to recreate these datasets assuming you own the software.

Getting FSL4 from a ZIP file

The whole dataset can be downloaded from here. It contains the audios for all sounds in their original format plus the metadata (including BPM annotations) for each file. Once downloaded you can use the convert_audio_files_to_wav.py script to convert all sounds to 44.1 kHz, mono, 16 bit PCM format:

python scripts/convert_audio_files_to_wav.py SOURCE_FOLDER --outdir OUTPUT_FOLDER

This conversion uses ffmpeg command line utility, therefore you'll need to have it installed in your system.

Getting FSL4 by downloading content from Freesound

Alternatively, FSL4 can be recreated running a script that downloads all audio files and metadata directly from Freesound. Be aware that using this script the resulting dataset might not be exactly the same as the one we used in our experiments as there exists the possibility that some files are missing or their descriptions have been altered.

The script is named create_freesound_loops_dataset.py and can be found in the scripts folder. Before running the script you'll have to set the FREESOUND_API_KEY and FREESOUND_ACCESS_TOKEN parameters found in the first lines of the script. To do that you'll need to create a Freesound account and follow these steps:

    1. Apply for Freesound APIv2 credentials accessing this url. The credentials are given instantly upon filling the form.
    1. Fill FREESOUND_API_KEY with the alphanumeric string of the Client secret/Api key column that will be shown after filling the form in step 1.
    1. Copy the Client id alphanumeric string (shorter than Client secret/Api key) that is also given after filling the form.
    1. Go to the following url:
https://www.freesound.org/apiv2/oauth2/authorize/?client_id=YOUR_GIVEN_CLIENT_ID&response_type=code

(replace YOUR_GIVEN_CLIENT_ID with the contents of Client id from step 3) and click Authorize!. Introduce your Freesound account credentials and copy the authorization code that will be given to you.

    1. Run the following command replacing YOUR_GIVEN_CLIENT_ID (as in the previous step), YOUR_GIVEN_FREESOUND_APY_KEY with the contents of FREESOUND_API_KEY, and AUTHORIZATION_CODE with the code you got in step 4.
curl -X POST -d "client_id=YOUR_GIVEN_CLIENT_ID&client_secret=YOUR_GIVEN_FREESOUND_APY_KEY&grant_type=authorization_code&code=AUTHORIZATION_CODE" https://www.freesound.org/apiv2/oauth2/access_token/

This will return something similar to this:

{"access_token": "e22ce72c0b18bf76a7da9d41ad1d7e146f3af309", "token_type": "Bearer", "expires_in": 86399, "refresh_token": "31cbd0fd394f7381a70b73befd4b97e7a9afdad9", "scope": "read write read+write"}
    1. Fill FREESOUND_ACCESS_TOKEN with the contents of the alphanumeric string access_token of the response.

By following this process you'll get an access token that will allow to download freesound audio files in their original format. This access token will be valid for 24 hours. If you run the script when the access token has expired, you'll need to follow again these steps to get a new token.

Once you have filled FREESOUND_API_KEY and FREESOUND_ACCESS_TOKEN you can now run the script to recreate the FSL4 dataset.

python scripts/create_freesound_loops_dataset.py --max_sounds 4000 path_where_you_want_to_store_the_dataset

Note that it might take a while to download all the sounds and metadata. Note also that because of throttling limits of the Freesound API the script might not be able to download all the sounds in one day. By default Freesound API allows 2000 requests per day, meaning that no more than 2000 sounds can be downloaded per day in their original quality. The solutions around that are either to run the script in consecutive days (sounds already downloaded won't be downloaded again) or to ask for more permissive limits to the Freesound API maintainers (see API documentation). Once sounds are downloaded the script will also convert the audio files to 44.1 kHz, mono, 16 bit PCM format. This conversion uses ffmpeg command line utility, therefore you'll need to have it installed in your system.

Getting the other datasets

Although we can not release the other datasets, APPL and MIXL can be rather easily reproduced if owning Apple's Logic Pro and/or Mixcraft 7 music production software music production software. If that is the case, the following instructions should help you getting the dataset ready.

APPL

Apple's Logic Pro comes with a number of audio loops bundled with the software. These loops are in .caf format and typically come with metadata embedded. To create APPL dataset you can follow these steps:

    1. Install Logic Pro (only available for OSX)
    1. Create a folder for the db in your data directory
    1. Extract the metadata from the .caf files using the software hosted in this public repository. Apple loops files are typically located in /Library/Audio/Apple Loops/Apple/.
    1. Save the .json file created for each loop in a folder called metadata inside the dataset folder.
    1. Run the script create_apple_loops_dataset.py that you'll find in the scripts folder. You'll probably have to tweak some bits of the script to provide the full path to the original audio files.
python scripts/create_apple_loops_dataset.py created_db_folder_in_step_1

The script will take care of creating 44.1 kHz, mono, 16 bit PCM format versions of the files and save them into the audio/wav folder of the dataset. After running it, the dataset should be ready for use.

MIXL

Similarly to Logic Pro, Mixcraft 7 also comes with a number of loops bundled. To create the MIXL dataset you can follow these steps:

    1. Install and open Mixcraft 7 (only available for Windows)
    1. Create an empty project
    1. Click on the Library tab at the bottom
    1. Select the Loops library and wait untill all sounds are imported/downloaded. When finished you can close Mixcraft.
    1. Create a folder for the db in your data directory and inside a folder create audio/original subfolders.
    1. Copy the contents of Mixcraft loops folder in audio/original. Mixcraft loops folder should be found in C:\ProgramData\Acoustica\Mixcraft\loops\.
    1. Locate the file Loops.mldb7 (also in the same Mixcraft loops folder) and copy it to the root db folder. Rename it to original_metadata.csv.
    1. Run the script create_mixcraft_loops_db.py that you'll find in the scripts folder.
python scripts/create_mixcraft_loops_db.py created_db_folder_in_step_1

The script will take care of creating 44.1 kHz, mono, 16 bit PCM format versions of the files and save them into the audio/wav folder of the dataset. After running it, the dataset should be ready for use.