Musicbrainz has used several audio fingerprinting systems over its lifetime.
All of them (so far) work in essentially the same way. It is generally a two-step process of submission and lookup. First, the raw audio is used to create a fingerprint, which is then submitted to a third-party server. This server analyzes the fingerprint, compares it to other fingerprints, and decides whether it is sufficiently different from known fingerprints as to issue a new ID.
Once this step is done, a fingerprint can be calculated for any file and this can be used to look up the corresponding ID.
This ID is associated with a given track (pre-NGS) or recording (post-NGS), and metadata can be gathered from there.
TRM (TRM Recognizes Music) IDs were MusicBrainz’ first audio fingerprinting system. This system was created by Relatable, and added to Musicbrainz in 2000.
This system worked reasonably well for finding duplicate music files on a local system, but had problems with collisions (different-sounding audio which got the same ID) and duplicates (same-sounding audio which has different IDs). The server was also incapable of handling the number of TRMs needed for Musicbrainz, and Relatable didn’t seem to be interested in supporting it further.
PUIDs are Musicbrainz’ second audio fingerprinting system. This was initially operated by formerly MusicIP/MusicMagic/Predixis, and bought by Gracenote in June 2011. Gracenote is expected to discontinue the public service soon, and it already appears to be largely non-functional.
This system was better than TRM, but still has several major problems:
- The fingerprint submission system is not open source, and as such cannot be included in Picard. (the lookup system is, and was included in Picard as of version 0.7.0, released in July 2006.)
- The fingerprinting process is slow, both on the client side and the server side
- Over time, the operators have become less and less interested in running the server, to the point where today it is barely working, if it works at all.
PUIDs are also quite opaque, being nothing more than a unique number referencing a database outside of MusicBrainz’ control. If/when that database goes away, they become useless. See their patent application for details on the technology. The client-side audio fingerprinting library is open-source.
It has several immediate advantages:
- It is open source.
- It is actively developed, along with supporting software.
- It gives the ability to visually compare music
- AcoustID fingerprints have their duration recorded, making it easy to discard certain incorrect links between recordings and acoustIDs.
Other audio fingerprinting systems
- The fingerprint in Kurt Rosenfeld's FDMF.
- MusicURI, part of the Mpeg-7 Audio DB project.
- jHears is an acoustic fingerprinting framework based on FutureProofFingerPrint design by Geoff Schmidt (formerly of Tuneprint). jHears is developed by Juha Heljoranta.
- AudioScout. Based on the pHash audio fingerprinting library, developed by the same authors. Uses the "Philips Robust Hashing" algorithm.
- Echoprint. Audio fingerprinting solution developed by the Echo Nest.
- [defunct] libFooID. An audio fingerprinting library used by, and developed for foosic.
- [defunct] Freetantrum. It seems to be a dead project (its home page was replaced with an advert for unrelated things in 2001), but it may be worth investigating and resurrecting the code they produced.
- Last.fm. The client-side audio fingerprinting library (fplib) is open-source. Uses the "Computer Vision for Music Identification" algorithm.
- [defunct] TRM, see above.
- [defunct] PUID, see above.
- Audible Magic. This article compares business models for Philips & Audible Magic
- AudioID. Developed at the Fraunhofer Institute for Integrated Circuits IIS and the Fraunhofer Institute for Digital Media Technology IDMT, now provided by mufin GmbH.
- Shazam. Proprietary music fingerprinting system, currently offered directly to users over the telephone.
- Rovi Media Recognition Service (formelly AMG LASSO).
- Gracenote MusicID.
- Philips. One of the first scalable audio fingerprint algorithms. Described here and here.
- And possibly most ridiculously of all: The Song Tapper