The previous post about beta testing our new fingerprint technology generated quite a bit of feedback – thanks to everyone involved.
Quick Update
- 10 million fingerprint submissions received so far (keep ’em coming)
- Lots of useful feedback and thoughtful questions
- We are frantically working on the server architecture in order to get a public-facing lookup service ready as soon as possible.
- Dedicated hardware for this is scheduled to arrive on Friday afternoon. Laurie will probably post pictures once all the LEDs are working.
Gory details
This weekend we started feeding the fingerprints en-mass into the fingerprint indexing service, which will tell us how many duplicate tracks there are and so on. This is in full swing, but we’ve only imported about 3m so far. A graph showing total fingerprints received, and total unique tracks is available here (updates twice an hour) – bear in mind that we have not imported all the fingerprints yet, so expect the ratio of ‘unique tracks’ / ‘fingerprints received’ to increase over the course of the week.
Although far from complete, we have the capability to query the currently imported data and return an internal unique ID, and the list of spellings used for that fingerprint:
$ ./query ../mp3/Pink\ Floyd/The\ wall\ \(disc\ 1\)/03\ Another\ brick\ in\ the\ wall\ \(part\ 1\).mp3
Fingerprint ID: 527153
--
Pink Floyd - Another Brick in the Wall (Part 1)
Pink Floyd - Another Brick In The Wall (Part I)
Pink Floyd - Another Brick In The Wall Part 1
Pink Floyd - Another Brick in the Wall, Par
Pink Floyd - Another Brick in the Wall, Pt. 1
Pink Floyd - The Wall (PartI)
Pink Floyd - Another brick in the wall (p..
Unknown - unknown
Or another example:
$ ./query ../mp3/Guns\ N\'\ Roses/Use\ Your\ Illusion\ II/04\ -\ Knockin\'\ On\ Heaven\'s\ Door.mp3
Fingerprint ID: 1211395
--
Guns N' Roses - Knockin' on Heaven's Door
Guns N' Roses - Knockin´ On Heaven´s Door
Guns N' Roses - Knockin On Heavens Door
I’m expecting a lot more versions of this one once we’ve cleared the fingerprint backlog… all those apostrophes ;)
What this shows is simply the list of known spellings for songs that sound exactly the same as the MP3s used as a test. The list of known spellings will increase as we import the rest of the data. It’s going to be trivial for us to order this list by popularity, and assume that the most popular spelling is usually correct. We are however, painfully aware that this is not going to be reliable enough in many cases.
The crux of the moderation task will be taking these sets of spellings (once album and popularity data is added) and distilling it into a single correct version
Once we have processed the fingerprint backlog (and hopefully updated the client to address some known issues) we will be in a better position to figure out how we’re going to do this.
Initial questions we hope to answer:
- How accurate will ‘the most popular spelling is correct’ assumption be?
- How many unique recordings make up the ~10m fingerprint submissions?
- Can we work with MusicBrainz for voting, and to publish a “common misspellings” dataset?
- Just how many ways to write “Guns N’ Roses – Knockin’ on Heaven’s Door” are there?
Comments
Paul
10 September, 17:25
And hopefully:
5) How can we resolve tracks to MusicBrainz IDs?
Moshroum
10 September, 17:25
What about ogg vorbis support for the fingerprinter? Its really easy to add support to the application with libvorbis and libogg.
Will MusicBrainz-Informations in songs used (for example to link to musicbrainz database)?
Rich
10 September, 17:27
Glad to hear an update – think this will drastically improve the quality of the last.fm data set!
Why not just assume the most popular spelling, and let users throw up any exceptions?
If more than 5 (say) people say it’s wrong, throw the top 10 alternatives up for a vote.
Thing is we don’t really want this hanging over last.fm for years…
RJ
10 September, 17:35
Paul: once we have the correct spellings, resolving to MBIDs is trivial with a direct text match. MBIDs will be included in the webservices like they are now for artists.
It’s even easier if we can tie the voting in to MB to start with tho. Open to suggestions here, we’re focussing on the architecture at the moment, and have no concrete plans for how final voting would work yet.
Moshroum: Yup, we plan to add vorbis and flac support amongst others
coxy
10 September, 19:17
Ace. It’s nice to see a good old graph. Everyone likes a good graph / chart.
I’m just nearing 10,000 tracks fingerprinted. I might eat a chocolate biscuit to celebrate.
I even blogged about the Fingerprinter; how’s that for loyalty. (though I still quite like MusicBrainz – as seems to be the general consensus).
Laurie
10 September, 20:16
I like graphs too coxy. That’s why I made it. I make hundreds. It’s all I do. As well as play with flashy LEDs.
Veiona
10 September, 20:18
Thanks for the Update!
mosquitokillah
10 September, 20:35
Thanx too for that great update!
UW
10 September, 20:40
I’m quite sure that the “the most common spelling is correct” approach will generate LOTS of wrong spellings!!
You often get wrong spellings when you insert a CD and let iTunes (or anything else) fetch the data from Gracenote. And many people don’t check if the information is correct or not.
Wrong spellings spread like a plague on the internet as many people don’t really care if their titles are spelled correctly or not. So if you go for the most common spelling, you’re in trouble! :-(((
BTW: There are even misspellings on tracks bought on download stores or downloaded from the artists’ websites.
Carey
10 September, 20:43
RJ: A direct text match won’t always yield the correct MB track ID. MB strives to catalog all the various releases of albums and tracks over time, and as I’m sure you know, there are plenty of cases where the “same” track is released again, perhaps even on the “same” album, but is different enough that it could have a different fingerprint. This is a sticky problem to which I’m not sure there is an easy solution.
B
10 September, 20:53
Why not use MB’s standards for determining which spelling to use? Users have put countless hours and discussions into correct formatting over there, and there’s no reason Last.fm should have to duplicate that effort. I echo the above commenter about how the most popular spelling is hardly guaranteed to be the most correct.
Mika
10 September, 21:19
Aww, no leadercharts.
Rob Szarka
10 September, 22:02
I agree: the intersection of “most popular” and “correct” isn’t going to be as large as you hope. The damn record labels aren’t smart enough to care, much less the average user…
I’m still curious whether your fingerprinter is preferring ID3v2 over ID3v1, or perhaps submitting both? It will make a huge difference for most of the 50,000ish tracks I’ve submitted. (Though I also made sure to send you some fingerprints for tracks I haven’t cleaned up yet, too.)
Thanks for the update, tho. :)
davekg
11 September, 01:06
Excuse my ignorance if this question is utter bunk: Why not have the client submit musicbrainz id’s (if available in tags) when it scrobbles and ignore any other tags or fingerprints altogether? This seems the easiest way to unambiguously identify an artist/release/track. (Of course, the other methods are still necessary if the MB tags are not present.)
Carey
11 September, 01:29
davekg: That’s is not a bad idea, unfortunately the fingerprinting application doesn’t seem to submit MB trackid, even if it’s present. Looking at the network traffic, I see a “&mbid=” parameter being submitted, but the value is always empty (even though all my tracks have MB trackid). This seems like a bit of a missed opportunity, but with 10M tracks already submitted, I guess there’s not a whole lot that can be done about it now.
Screwtape
11 September, 05:12
As others have said – thanks for keeping us up-to-date on the project!
I think I’ve found a bug with the finger-printer, though – every time it tries to finger-print a track that’s more than an hour long, it dies with a network error (yes, I have a few tracks over an hour long in my collection; DJ sets and recordings of radio shows). I wouldn’t mind if it skipped finger-printing tracks over an hour long, but every time I say ‘fingerprint my music folder’ it resumes at the long track and predictably fails again. I’m now working through individual artist folders, but clicking a few hundred tick-boxes is not nearly as friendly as just clicking one.
Philip Jägenstedt
11 September, 05:58
It’s very nice to see that the MusicBrainz issue is being taken seriously, I hope there can be some reasonable solution. Since not all the music on the planet is in MusicBrainz, this could potentially be a tool for discovering popular music which is not in the MB database and adding it.
Asbjørn Ulsberg
11 September, 11:04
Very interesting to read how the fingerprinting business is going! I have one question, however: Will the fingerprint of two files of the same song with a different length, different bitrate and perhaps even different recording medium be the same?
Will for instance different live recordings of a given song have the same fingerprint? Will those live recordings share the fingerprint with the studio recording? What about different studio recordings of the same song? And what if the same recording of the same song is ripped from different media (CD and LP as a very hypothetical example)?
How good is the fingerprinter to catch variations like this and how does it classify them? Some variations obviously separates one song from another (live vs. studio for example) while other variations are bi-effects of the ripping procedure (different length, different media, etc) and the two files should be considered the same song and thus share fingerprint.
John
11 September, 11:45
Screwtape, the same is happening me.
All my long DJ set recordings (much of them over an hour long) are getting a network error.
Even more annoying is that the software then stops fingerprinting. I tried to leave it on all night last night, but it stopped after half an hour because of this. It should keep going after getting an error, cache the results, then submit them later.
john
11 September, 11:50
One thought, it’s probable that the kind of people who download this and run it, are the kind of people who have their tags spelled correctly anyway…
Moshroum
11 September, 12:03
john, this thing will be integrated in the normal client.
Moshroum
11 September, 12:06
@RJ: Good to hear that their will be ogg vorbis support. But whats about reading mbids from files to get a better idea what is the correct musicbrainz entry (text matching will not work when many people tag their songs wrong).
Erik
11 September, 12:10
Asbjørn: our fingerprints distinguish between different recordings while remaining invariant to noise differences that result from different media.
RJ
11 September, 12:23
I didn’t realise until it was too late that the FP app wasn’t snagging MBIDs (the field was there but is always blank, doh). Anyway, the best outcome for us is to distill the sets of spellings into MBIDs – then we can leave the unenviable task of selecting the most appropriate metadata up to the musicbrainz folks (many of which use lastfm too of course).
Another outcome as Philip mentioned is that we should be in a position to name the most popular content we are seeing that is not in MB, and feed that back to them.
Paul
11 September, 14:19
For those keeping track of the answer to the question: Just how many ways to write “Guns N’ Roses – Knockin’ on Heaven’s Door” are there? RJ has posted the 413 variations of ‘Guns N’ Roses’ that the last.fm fingerprinter has encountered so far in a comment on blog post at:
http://blogs.sun.com/plamere/entry/how_many_ways_to_spell#comments
Marcel
11 September, 14:55
IMO you should use MusicBrainz database for determining which spelling to use. The popularity approach seems very inaccurate to me.
(sorry if it has been mentioned in the previous comments, I only read a few ones)
Phil
11 September, 18:54
When it comes to electronic music of all kinds, I’d prefer discogs, I mean, their entire site is about “how to spell this song the correct way”…same goes for albums, Eps, releases of all kinds.
UW
11 September, 20:43
I second what Phil said: Working together with Discogs would be great!! They really try to get spellings right!
I would be very neat to get some interconnection between last-musicbrainz-discogs going!
BTW: Thanks for all the work you’re doing!
PAStheLoD
11 September, 22:07
Thank you for the update, very much appreciated.
What about similar tracks? I suppose the hashing algorithm used by the fingerprinter is a “small input change, big output change” kind of function, then maybe using track names and other data to link different IDs together? Then you can really get unique results.
Pla
11 September, 23:14
What about reading and sending MusicBrainz tags along with the fingerprints if they are available?
Picard stores a lot of useful information in those tags
for example:
G M
12 September, 01:21
Please release a version of the fingerprinter for the Linux command line.
Ian Smith
12 September, 01:52
I’ve uploaded 1781 so far– but man, every two songs I have to click the error message. What I’ve been doing is setting it up with a script so my cursor jumps to the button every time it comes up. Luckily I’m not easily annoyed, or I’d‘a given up. Can’t wait ‘til those dedicated servers come in on Friday…
Zeta
12 September, 08:20
The million dollar question: Are the wrongly merged artists from the previous moderation system gonna be “demerged”?
Mark
12 September, 08:39
Is the plan here to automatically correct incoming submissions for display or purely for background processing only? While changing for display could be a “good thing”, it is also a “bad thing”… A particular example I would use is live tracks… on my charts I like to distinguish live tracks I scrobble from their album counterparts and so I change all my live titles to this format
<title> <album> [live]
So for instance, Comfortably Numb on the Pulse album would be scrobbled as
Comfortably Numb (Pulse) [live]
This allows people looking at my charts to see that exactly which version of Comfortably Numb is in my charts.
Background processing the Pulse version of Comfortably Numb above with other users listening to the same track is clearly a good idea, but automatically changing the display to just
Comfortably Numb
will either mean I have multiple Comfortably Numb entries in my charts with no idea which is which version, or you will pile all versions into one which is just plain wrong, IMO.
So, I hope you are only doing corrections in the background and leaving individual user displays alone??
While I appreciate LastFM is not my personal chart generation system, I do feel that merging all versions of the same song into one is fundamentally wrong and could skew results by inflating certain titles results due to all the various versions of a song that float around.
Cheers,
Mark
Neueklick
12 September, 08:55
i really would like to help but the servers are constantly »temporary« overloaded …
Greg_x7
12 September, 09:31
Yeah, I have the same problem Neueklick :/
Julian
12 September, 09:35
I think the fingerprinter can distinguish between studio and live versions of the same song, so your song titles will be safe ;)
Onno
12 September, 09:59
If I understand correctly the fingerprinter listens to the songs and determines which songs ‘sound’ alike. So a live version will be treated different then the studio version.
Even more, it would mean that different live version will be treated different, since they most likely will not sound the same.
warp
12 September, 12:41
Why not submit the musicbrainz ids along with the other metadata when a fingerprint is submitted? The most common mbid for a fingerprint is much more likely to have the ‘correct’ version of track name than the most common track name. right?
RJ
12 September, 13:14
A linux command-line client is on the way, that will also send MBIDs correctly.
AdriaNnLA
12 September, 23:52
Concur with warp’s observation (in addition to several of those above), in that…
whilst ripping CDs, I often have to correct the data that comes in from each of the magical sources, and
99.9% of tracks I’ve ever downloaded, whether from a p2p (especially), artists site, labels site, amazon, emusic, last, magnatunes, mp3.com etc have errors of one sort or another within the tags, so
simply giving credit to the most popular spelling is more than likely to not result in the most accurate, and
while tracks containing MBIDs (whether TRM or PUID) are not necessarily guaranteed (but certainly more likely to be accurate) extreme weight should probably be given where they exist. The more extreme of course given to PUIDs.
(i.e. PUID > TRM > ‘no MBID’)
I’d guess that close to 90% of my collection has been through Picard and contains PUIDs; so if you didn’t collect them the first go ‘round, I’ll be happy to clear and fingerprint again if/when a new version allows for such.
Lovell
13 September, 08:00
I’m glad to see things going so well.
As far as the finding the “correct” spelling, I definitely see where the difficulty will be, as the problem is that you want to find a way to do it so that you don’t need human interaction (i.e. whatever computer is doing the analysis can automatically determine which way the “correct” one). The idea of using the one most common on songs with an MBID is an interesting idea, but you’d have to have a bunch of people resubmit there tracks because the client isn’t taking MBIDs right now. I’m more optimistic than most about the most popular version of a song spelling being likely to be correct (especially with songs that will have thousands or even millions of copies), but I’m just generally excited to see what you guys come up with. Keep up the good work!
Fred
13 September, 14:25
Just to say that I consider this effort really needed and it will improve even more the last.fm service.
One of the bigger problems I see now is the proliferation of different names for a song. For instance the song Ecstasy of Gold (in italian L’Estasi Dell’Oro) by Ennio Morricone has the following popular name versions.
The Ecstasy of Gold
The Ecstacy of Gold
L’Estasi Dell’oro (The Ecstasy Of Gold)
L’Estasi Dell’Oro
I consider really important being able to unify these various versions into one entity.
Other question/suggestion: I think the last.fm client application should fingerprint tracks and if the user wishes so, he could approve the corrections/suggestions to make to the mp3 tags. Is this what you guys are planning to do?
Xavier
13 September, 16:00
Zeta’s question is the good one imho
Dirk
13 September, 19:08
That would be awesome. Though a collaboration between discogs, last.fm and MusicBrainz (as UW mentions it) would be even better :)
Adriaan
14 September, 14:00
I love you Last.fm. To celebrate I’ve written you a Wikipedia article: http://af.wikipedia.org/wiki/Last.fm
Tobbyno
14 September, 14:06
Could this “fingerprint” thing also be applied on tags?
I mean to merge pop-rock, pop rock, pop/rock all in the same tag (pop/rock).
Or Post-Hardcore, post-HC, post hardcore…
hope you get what i’m trying to explain, meh my english sux :D
Adriaan
14 September, 14:12
I don’t like forward slashes in my tags! And some people do like it, so don’t merge them! Except if every Last.fm member agrees that they don’t mind not having slashes in their tags…
Onno
14 September, 14:35
Tagging is based on individual preferences, so tags should be left alone.
Simon
14 September, 15:36
Does this mean that all the clients will soon submit MBIDs correctly?
Amr Hassan
15 September, 08:00
if another client that submits MBIDs and has caching features is released, i’d gladly re-fingerprint all of my collection again..
the real pain the first time was the overloaded message.
Adriaan
15 September, 09:59
Hello people of Last.fm. Can you please add a statistic to our profiles that says “Tracks fingerprinted: xxxx”? It’s all about statistics, isn’t it? This is also a good way of displaying how many files we have in our library…
john
15 September, 11:05
Yeah I’d like if there was a record being kept of our fingerprinting work! :D
Sceptic
15 September, 12:02
I’d hope such a feature would be made optional. I have no interest in showing the world how large my collection is.
Adriaan
15 September, 14:39
Yes, make it optional then. However, does Sceptic want to hide all his charts too? I don’t see why you want to hide how many fingerprints you submitted.
Zeta
16 September, 09:40
I assume they are, but I’ll ask again: are the wrongly merged artists from the previous moderation system gonna be “demerged”?
Sceptic
16 September, 14:58
As for my charts, I mostly stream music while at work, so that’s hardly a good way to determine what’s on my harddrives.
Den_Dolf
16 September, 15:03
some improvements on the fingerprinter would be nice, like the ability to make it read my music from the network storage, and to have this tool completely silent, only in the system tray :)
Amr Hassan
16 September, 15:42
i’ve noticed that since like last thursday the number of unique tracks fingerprinted in the graph hasn’t increased significantly. could that be all the music in the world already!
DiamondDave
16 September, 18:01
I’ve been getting troubles with the network disconnecting too, seems the fingerprinter runs into trouble with longer tracks on slower machines.
jhimbob
17 September, 12:55
i was just confused that most people’s music is the bane of bad tagging, but then when i listened to last.fm radio, there were lots of mistakes as well. is that a database problem or the mistake of the person that uploaded it?
AKMA
17 September, 13:21
My obsession extends to ascribing correct dates to releases, a problem that re-releases and compilation albums render problematic. Any help for us chronologists?
Jester-NL
18 September, 09:42
I notice in the graph that the amount of unique tracks has been stuck for a few days now.
Is that because the current amount of fingerprinted tracks is now only consisting of ‘doubles’ or is the processing halted?
Adriaan
19 September, 19:48
Hello! When will the chart be up? Are you still working on this?
RJ
20 September, 11:55
Yep, we’re re-inserting because of some nice speed improvements on the FP server – at this rate the network round-trip time is going to be by far the limiting factor when doing FP lookups :)
Adriaan
20 September, 13:06
Wow cool! You are the best!
xD
Adriaan
20 September, 14:39
Hi RJ, I know this is not the place, but could you please inform me whether you guys are planning in giving us more stats in that section under our avatars? Like Unique artists, unique albums, unique tracks, tracks fingerprinted, tracks per day/month/year, shoutbox comments, etc. I know some of it might be difficult to implement, but I mean, will you consider it in the future and so on? And if you do consider it, will you make it optional? I think there are somewhat people wondering about this…
Anyway, thanks, I am enjoying your site.
Daryl
20 September, 16:29
Any ETA on the Linux command line fingerprinter?
Amr Hassan
21 September, 00:36
any ETA on anything? we’re in desperate need of new updates. you got us all excited for a while now :D
Thomas
21 September, 08:20
I second that. This sounds incredibly exciting and I would really like to see some more information on the on-going proccess.
Keep up the great, innovative work!
Ariel
21 September, 15:16
Is there a way to use this fingerprinter without having to see and watch it, cos it stops every once in a while, way too soon? I mean, I can only use this with a few folders at a time, and I have THOUSANDS of folders to fingerprint.. it’s like, a pain in the……..
Ariel
29 September, 05:50
Is there a way to fingerprint more than 5 songs at a time!? It’s annoying and frustrating..
James
11 October, 21:00
Why can’t you fingerprint .wmas? They make up most of my collection.
Ville
12 October, 10:53
@James: There is no WMA support because it is closed format & there is no open source libraries that could read it, I suppose. (And last.fm clients must work also with other OS’es than Windows)
nakke
21 October, 16:14
RE: merging tags
I think at least obvious stuff like “post rock”, “post-rock” and “postrock” should all be merged to the same tag.
Comments are closed for this entry.