Download datasets

Data size — 2022/02
Audios files	800,000+
Average size	100kB
Total size (est.)	80GB

Find your target

Categories on Wikimedia Commons are organized as follow:

• Commons:Category:Lingua Libre pronunciation by user
• Commons:Category:Lingua Libre pronunciation (by language)

Download datasets via click

Download by language:

1. Open https://lingualibre.org/datasets/
2. Find your language, naming convention is: {qId}-{iso639-3}-{language_English_name}.zip
3. Click to download
4. On your device, unzip.

Post-processing
Refer to the relevant tutorials in #See also to mass rename, mass convert or mass denoise your downloaded audios.

Programmatic tools

The tools below first fetch from one or several Wikimedia Commons categories the list of audio files within them. Some of them allow to filter that list further to focus a single speaker, either by editing their code or by post-processing of the resulting .csv list of audio files. The listed targets are then downloaded at a speed of 500 to 15,000 per hours. Items already present locally and matching the latest Commons version are generally not re-downloaded.

Python (current)

Dependencies: Python 3.6+

Petscan and Wikiget allows to download about 15,000 audio files per hour.

1. Select your category : see Category:Lingua Libre pronunciation and Category:Lingua Libre pronunciation by user, then find your target category,
2. List target files with Petscan : Given a target category on Commons, provides list of target files. Example.
3. Download target files with Wikiget : downloads targets files.

Comments:

• Successful on November 2021, with 730,000 audio downloaded in 20 hours. Sustained average speed : 10 downloads/sec.
• Some delete files on Commons may cause Wikiget to return an error and pause. The script has to be resumed manually. Occurrence have been reported to be around 1/30,000 files. Fix is underway, support the request on github.
• WikiGet therefore requires a volunteer to supervise the script while running.
• As of December 2021, WikiGet does not support multi-thread downloads. Therefore, to increase the efficiency of the download process it is recommended to run the Python Script on 20-30 terminal windows simultaneously. Each terminal running WikiGet would consume an average of 20 Kb/s.
• WikiGet requires an stable internet connection. Any disruption of 1 second would stop the download process and it requires manual restart of the Python Script.
• Manual for PetScan
• Any question about downloading datasets can be made on the Discord Server of Lingua Libre : https://discord.gg/2WECKUHj

NodeJS (soon)

Dependencies: git, nodejs, npm.

A WikiapiJS script allows to download target category's files, or a root category, its subcategories and contained files. Downloads about 1,400 audio files per hour.

1. WikiapiJS is the NodeJS / NPM package allowing scripted API calls upon Wikimedia Commons and LinguaLibre.
2. Specific script used to do a given task:
   • Given a category, download all files : https://github.com/hugolpz/WikiapiJS-Eggs/blob/main/wiki-download-many.js
   • Given a root category, list subcategories, download all files: https://github.com/hugolpz/WikiapiJS-Eggs/blob/main/wiki-download_by_root_category-many.js

Comments, as of December 2021:

• Successful on December 2021, with 400 audios downloaded in 16 minutes. Sustained average speed : 0.4 downloads/sec.
• Successfully process single category's files.
• Successfully process root category and subcategories' files, generating ./isocode/ folders.
• Scalability tests for resilience with high amounts requests >500 to 100,000 items is required.
• Performance improvements are under consideration on github.

Python (slow)

Dependencies: python.

CommonsDownloadTool.py is a python script which formerly created datasets for LinguaLibre. It can be hacked and tinkered to your needs. To download all datasets as zips :

• Download scripts :
  • create_datasets.sh - creates CommonsDownloadTool's commands.
  • CommonsDownloadTool/commons_download_tool.py - core script.
• Read them a bit, move them where they fit the best on you computer so they require the minimum of editing
• Edit as needed so the paths are correct, make it work.
• Run create_datasets.sh
• Check if the number of files in the downloaded zips matches the number of files in Commons:Category:Lingua Libre pronunciation

Comments:

• Last ran on February 2021, stopped due to slow speed.
• This script is slow and has been phased out as Lingualibre grown too much.
• The page may gain from some html and styling.
• Proposals go on https://phabricator.wikimedia.org/tag/lingua_libre/ or on the LinguaLibre:Chat room.

Java (not tested)

Dependencies:

sudo apt-get install default-jre    # install Java environment

Usage:

• Open GitHub Wiki-java-tools project page.
• Find the last Imker release.
• Download Imker_vxx.xx.xx.zip archive
• Extract the .zip file
• Run as follow :
  • On Windows : start the .exe file.
  • On Ubuntu, open shell then :

$java -jar imker-cli.jar -o ./myFolder/ -c 'CategoryName'     # Downloads all medias within Wikimedia Commons's category "CategoryName"

Comments :

• Not used yet by any LinguaLibre member. If you do, please share your experience of this tool.

Manual

Imker -- Wikimedia Commons batch downloading tool.

Usage: java -jar imker-cli.jar [options]
  Options:
    --category, -c
       Use the specified Wiki category as download source.
    --domain, -d
       Wiki domain to fetch from
       Default: commons.wikimedia.org
    --file, -f
       Use the specified local file as download source.
  * --outfolder, -o
       The output folder.
    --page, -p
       Use the specified Wiki page as download source.

The download source must be ONE of the following:
 ↳ A Wiki category (Example: --category="Denver, Colorado")
 ↳ A Wiki page (Example: --page="Sandboarding")
 ↳ A local file (Example: --file="Documents/files.txt"; One filename per line!)

See also

• Help:Renaming
• Help:Converting audios
• Help:Embed audio in HTML
• phabricator:T261519