The user can paste a youtube url which will then be analysed, fingerprinted and matched against a database of 7+ million audio fingerprints. It does not only identify a single song but is able to identify multiple songs contained in a single file or video and generates a timeline listing which tracks it contains at which time.
Our matching algorithm is based on the open source echoprint-codegen fingerprinting method, which we have built our own stack around:
- Replaced Solr/Tokyo Tyrant with Elasticsearch
- Reimplemented matching-logic
- Crawlers search multiple sources for audio files to be indexed (mp3s arent stored long term, only fingerprinted then deleted)
- Indexing about 1 new track per second
- Found method to verify unrealiable ID3 tags (in progress, current database also includes unferified)
Does it include music from oft-used music sources in YouTube videos such as AudioMicro and Incompetech[1]? I'd guess it's not really possible on the AudioMicro front as they're paid-for music that'd cost you a fortune to index but may be worth adding the latter
First we rewrote echonests truescore logic in perl and then altered slightly and implemented some extra checks to further try to exclude false positives. We also believe what they used in the late song/identify API might have been different from what is open sourced in https://github.com/echonest/echoprint-server
Also we pack each individual hash before storing in Elasticsearch and gained at least 50% storage space this way.
Our Fingerprint data is quite different from theirs(unreliable ID3 tags, N versions of same track) which is why we needed some tweaks.
So far the matching is still far from perfect...
Whether we will open source the whole thing at some point we don't know yet.
When you say the matching is far from perfect, is that at your end or on the part of the echoprint / echonest code? You made tweaks because you found issues with what they were doing....?
The reason for it being far from perfect is likely a combination of both. If the correct song is indexed there is a high probabiliy for us to find the right match. However if its not, with a bit of bad luck a false positive can happen easily with the default solution (and ours too).
Also when analysing a youtube video it can happen that in a 30sec snippet only 10 secs are a matching song and 20 are unrealated or 15 are one matching song the other 15 match a different one in which case 2 tracks or multiple versions of 2 different tracks will have relatively OK scores. Deciding what to consider a match (or whether to try different queries for the same or slightly altered timespan prior to deciding) is not trivial in these cases and our changes are mostly concerning when a match will be considered a match by altering thresholds and how matching truescores will be looked at in relation to other fingerprints true scores. Due to issues like these, specifying a timeframe for analysis will often produce better results.
song/identify supported both ENMFP and Echoprint, and AFAIK the Echoprint matching path was exactly the same as is published on Github.
I know at some point we did adapt the Solr end (for example, we removed the N most occurring codes) for speed optimizations.
Many users of Echoprint in the wild have adapted the python matching logic for their use case as well as changed the hash update rate on the codegen. A great modification to watch was Sunlight Labs' "Ad Hawk", which ID'd commercials: https://github.com/sunlightlabs/adhawk
Do you have any intuition for whether the echhoprint-codegen algorithm would be suitable for saying whether two voice recordings match? One would be a little lossy, the other pretty much perfect.
Echonest can work with voice but is optimized for music so you might encounter a lot of false positive with it.
Check out the echonest board on google. It's a recurring topic.
That's an absolutely creative way of doing it! Congrats.
I'm interested in how you did it. For example, how do you exactly "crawl the internet"? Did you have a bunch of sites that you've pre-selected and then just crawl those or did you actually follow through on links?
One-click hosters and cloud hosters with sharing options(such as docs.google.com), several music streaming sites, usenet.
Also the amount of mp3s hosted on regular webservers which are indexed by and easily found using the usual search engines is mindblowing
Music used in compilations, ads, intros etc are typical usecases.
Mixtapes & livesets of course are also great. But when pitch/bpms have been altered more than 1-2% we currently get a lot of false positives. We are still working on finding a solution to this.
I think I've read on the echonest board that the most common solution is to index multiple pitch variants of the same songs. Apparently that's what Shazam does.
Also the guys from Trax-air.com are doing something pretty similar to you guys but with pitch/bpm bending support.
So far we've been hesitant to go that route as the current 7m fingerprints come down to 42m+ documents in Lucene.
I.e. If we were to save 20 versions of each fingerprint in different bpms we'd quickly have about a billion+ documents on our hands to be searched on every query (with ~1000-2000 hashes per document)
For anyone often wondering about music in songs I will recommend Shazaams OSX app. It sits in the menu bar and listens to music and if it recognises something it will send a notification and add it to a list [0].
Watching youtube video, movies or just having someone else play something and it usually finds it without problems.
It's a different use case than OPs app though, which is more on demand I guess.
Shazam Mac osx app is quite powerful. It doesn't only listen to your mic so it will detect the song even when the sound is off.
Also it often detects a song playing in somebody's else headphones at our office.
Kinda creepy sometimes.
On a slight tangent: I'd love a client that could identify my MP3 collection, and rename it and retag it (under some kind of supervision). Ideally it'd do the dentification in a batchmode when it got Internet connectivity (but this is perhaps an unreasonable requirement). And to make it perfect it would let me listen to and delete tracks.
I have a huge unweildy collection of MP3s and I can't bring myself to just delete gigabytes of music.
Tried with a couple of videos, went out to grab lunch and 30 minutes later it is still stuck (with the progress bar extending to the first t in "http://"). Sorry :-/
I think what we need most is a community-backed source of fingerprints. Because the authority-based approach only works well for popular songs (at least, that is my impression, based on frequently using commercial recognizing apps).
This is very interesting - I see this as an ideal concept to be paired with one of the existing commercial solutions, e.g. Shazam.
If you're indexing all the random and free stuff out there, you're picking up a lot of material that may have never been commercially released or has not been re-released digitally. At the same time, Shazam, YouTube ContentID, Apple's iTunes Match, etc. have access to an extremely large set of references which (more than you could have) contain 99% accurate metadata. ContentID definitely picks up multiple songs in mixes, as well as pitch changes, with a high degree of accuracy (assuming the master sound recording has been submitted to YouTube).
A submission system would be great too, or some way for persons to tag stuff themselves ala discogs, etc.
45 comments
[ 4.4 ms ] story [ 97.1 ms ] threadOur matching algorithm is based on the open source echoprint-codegen fingerprinting method, which we have built our own stack around:
- Replaced Solr/Tokyo Tyrant with Elasticsearch
- Reimplemented matching-logic
- Crawlers search multiple sources for audio files to be indexed (mp3s arent stored long term, only fingerprinted then deleted)
- Indexing about 1 new track per second
- Found method to verify unrealiable ID3 tags (in progress, current database also includes unferified)
- mogilefs as primary data store for fingerprints
- perl everything
We also provide a free music identification API.
Any feedback would be much appreciated!
Thanks for clearing that up, good luck with your site!
[1] http://incompetech.com/music/royalty-free/
Do you plan to document/opensource you work ?
Also we pack each individual hash before storing in Elasticsearch and gained at least 50% storage space this way.
Our Fingerprint data is quite different from theirs(unreliable ID3 tags, N versions of same track) which is why we needed some tweaks. So far the matching is still far from perfect...
Whether we will open source the whole thing at some point we don't know yet.
http://static.echonest.com/echoprint_ismir.pdf
I know at some point we did adapt the Solr end (for example, we removed the N most occurring codes) for speed optimizations.
Many users of Echoprint in the wild have adapted the python matching logic for their use case as well as changed the hash update rate on the codegen. A great modification to watch was Sunlight Labs' "Ad Hawk", which ID'd commercials: https://github.com/sunlightlabs/adhawk
Thanks.
Did you just crawl random sites and search for .mp3 content on them? Or did you have a set of pre-defined search sites to craw?
Could you describe the usecases ? Is it for mixtapes uploaded on youtube by DJs, or Over-The-Air recognition in music festival videos ?
Because Music ( single tracks ) uploaded on youtube is usually already identified so it could be found.
Also the guys from Trax-air.com are doing something pretty similar to you guys but with pitch/bpm bending support.
Would be interested to know if you figured a way around that!
Watching youtube video, movies or just having someone else play something and it usually finds it without problems.
It's a different use case than OPs app though, which is more on demand I guess.
[0] http://i.imgur.com/0At4lJ6.png
http://www.cnet.com/how-to/siri-can-now-name-that-tune-via-i...
It works fairly well.
http://www.greenbot.com/article/2873722/how-to-perform-song-...
Again, works fairly well.
On a slight tangent: I'd love a client that could identify my MP3 collection, and rename it and retag it (under some kind of supervision). Ideally it'd do the dentification in a batchmode when it got Internet connectivity (but this is perhaps an unreasonable requirement). And to make it perfect it would let me listen to and delete tracks.
I have a huge unweildy collection of MP3s and I can't bring myself to just delete gigabytes of music.
I used it to tag a massive amount of partially labelled and mostly metadata-free music files some time ago and it worked a treat.
Gets flagged as copyrighted music even though there is no music?
So many YouTuber's would thank you for a service that did this.
If you're indexing all the random and free stuff out there, you're picking up a lot of material that may have never been commercially released or has not been re-released digitally. At the same time, Shazam, YouTube ContentID, Apple's iTunes Match, etc. have access to an extremely large set of references which (more than you could have) contain 99% accurate metadata. ContentID definitely picks up multiple songs in mixes, as well as pitch changes, with a high degree of accuracy (assuming the master sound recording has been submitted to YouTube).
A submission system would be great too, or some way for persons to tag stuff themselves ala discogs, etc.