|Resolved||kaldari||T120288 Enable MP3 uploads on Wikimedia Commons and TMH playback|
|Resolved||kaldari||T162395 Add .mp3 to the list of accepted file types on Wikimedia Commons uploads|
|Open||None||T134802 Improve the curator workflow for reviewing new files|
|Open||None||T120453 Copyvio tools for Commons|
|Open||None||T132650 Copyright detection (acoustic fingerprint matching) for audio files|
Not sure which projects this would fit into. It could be written as a bot or an extension. However, it would probably be wise to avoid WMF involvement lest Rights Holders point out policing to judges.
@zhuyifei1999 Most of the work for this task isn't in setup (a few afternoons), but in the ongoing maintenance (updating code/databases, adding new databases, reviewing results). So the overhead of running a GSoC isn't worth it.
Regarding images there are several commercial cloud API CV systems now available. Rate limited and will likely only talk to @wikimediafoundation.org email addresses. But that's for another task.
@Aklapper; This was originally proposed as a bot task in which I was interested in attempting. However, after other options proposed (extension), I am waiting to hear what others think the best option would be. Therefore it'd be inappropriate for me to claim this task without having the skills (if it were an extension) to complete it. If you disagree with the priority, by all means change it.
An extension is unlikely to be installed since it needs to be written by a WMF employee/contractor (some trust thing), would take years for approval, and WMF doesn't clean up messes (They got unpaid volunteers to do that). A bot that hits watchlists with potential speedy deletions is "good enough".
@Dispenser: That statement is incorrect. Steps and criteria for deploying extensions are available. For general meta-level discussions unrelated to this task's topic, please use better suited venues. Thanks.
There alot of Acoustic fingerprint services, so many we nuked our list on Wikipedia. ACRCloud has a rundown on their competitors, with some praise for FLOSS AcoustID. A 2012 Investigation picked Last.fm (90 million songs).
- AcoustID - Open Sourced. API Free for non-commercial and limited to 3 requests/sec. The database is CC-BY-SA 3.0 and PD and using MusicBrainz Picard and VLC player. Ubuntu package python-acoustid; Needs full clean source for matching. 8.3 or 25.5 million fingerprints.
- Gracenote SDK (GNSDK) - Free for non-commercial, limited to 100-1,000 API calls per day or less. 200 million tracks.
- Echo Nest Audio Fingerprinting - Open source (GitHub). Spotify bought Echo Nest in Mar-2014, the Song identification API was shutdown Jan-2015, Email & telephone are dead (May-2017), and Spotify has no Upload features. It covered 13 million songs (Public) and 30 million songs (proprietary).
- MooMa.sh released 7.8 million fingerprints (102 GB) before discontinuing the Public API
- ACRCloud Music Recognition - 14-day trial, Commercial (Pricing). Emailed, they're willing to work with us. 3rd party linking (for verification): Spotify, Apple Music, YouTube. GitHub: Audio recognition file scanner. 40 million fingerprints.
- Doreso - developer.doreso.com seems to be down since end 2015.
- Axwave - Corporate licensing only?
- Rovi Media Recognition - Tivo, so likely commercial only
- No APIs: Shazam or SoundHound
Implementation update. See T132650#3273150 for the survey of technology.
- AcoustID (Python lib): Integrated with an IRC bot for testing. Match count is low. There should be no obstacles bringing it to Tool Labs.
- Echoprint with MooMa.sh fingerprints: The 100 GB is non-trivial, but doable on Tool Labs
- ACRCloud: Signed up for trial, used sample code and the binary blob to test. Spoke with staff, willing to give access with our low usage in exchange mention on user page and blog post (I think theirs) on their technology (Like how we used it).
- Gracenote: Downloaded proprietary SDK, still need to compile it.
@kaldari I've experimentally added AcoustID into an IRC bot notifying Commons admins of audio and video files uploaded by newbies. Three files were recognized in the past week of WP0 abuse, I'd say it's <10% effective.
Could be a demographics thing, Moroccan Teenagers (WP0) vs Audiophiles (MusicBrainz Picard).