Well that explains it. I was on a support call to Dell the other day and was put on hold with a song playing in the background. When I eventually got to speak to a representative he told me the song name and artist performing it.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
SHAZAM..How the hell does it do that?
- Thread starter cenetti
- Start date
- Sort by reaction score
Well that explains it. I was on a support call to Dell the other day and was put on hold with a song playing in the background. When I eventually got to speak to a representative he told me the song name and artist performing it.
The recognition technology used by Shazam (www.shazam.com) is owned by Landmark Digital (www.landmarkdigital.com). It is a patented fingerprinting technology that recognizes practically any audio that is within the recognition database very quickly and with a very small amount of audio even under heavy noise. Shazam provides recognition services on phones and smartphones all over the world (and now the iPhone).
For the person who mentioned classical - classical is a huge library that is difficult to keep cataloged, which is likely the reason it wasn't picked up. Other genres should work very well.
//////////////////////
I visited the 'landmarkdigital.com' site; it looks like they use a simple neural network to do this pattern recognition stuff, essentially mimicking (presumably) how people do the same thing.
The network takes as inputs, certain song attributes (e.g., tempo, tone, instruments), and provides as output, its best guess at the song. A bunch of nodes are placed in between the inputs and output, and these nodes are trained (i.e., values changed) by running a bunch of known songs through the network and making small changes to the node's values until the network gives the right prediction for the song (i.e., gives the correct output). When you query the database, you're essentially providing an input to this trained network, which just spits back its output to you.
Presumably, this company is constantly retraining and updating their network by running new songs through the database, and updating the weights of the nodes accordingly (i.e., they make sure the network still predicts the correct output for new songs; and if it doesn't, they change the weights so it does).
I do wonder what inputs they're using for the network, and how your recording of the song gets converted to values for each of the inputs. I think this is where the majority of the cleverness in the technology resides. A bit of spectral analysis is probably done (FFTs), but other than that, I'm not too sure.
My best guess,
-Clayton
For the person who mentioned classical - classical is a huge library that is difficult to keep cataloged, which is likely the reason it wasn't picked up. Other genres should work very well.
//////////////////////
I visited the 'landmarkdigital.com' site; it looks like they use a simple neural network to do this pattern recognition stuff, essentially mimicking (presumably) how people do the same thing.
The network takes as inputs, certain song attributes (e.g., tempo, tone, instruments), and provides as output, its best guess at the song. A bunch of nodes are placed in between the inputs and output, and these nodes are trained (i.e., values changed) by running a bunch of known songs through the network and making small changes to the node's values until the network gives the right prediction for the song (i.e., gives the correct output). When you query the database, you're essentially providing an input to this trained network, which just spits back its output to you.
Presumably, this company is constantly retraining and updating their network by running new songs through the database, and updating the weights of the nodes accordingly (i.e., they make sure the network still predicts the correct output for new songs; and if it doesn't, they change the weights so it does).
I do wonder what inputs they're using for the network, and how your recording of the song gets converted to values for each of the inputs. I think this is where the majority of the cleverness in the technology resides. A bit of spectral analysis is probably done (FFTs), but other than that, I'm not too sure.
My best guess,
-Clayton
midomi is absolutely worthless. I remember trying and trying to get it to return ONE correct song name/info and im a musician, so I have a good sense of pitch. I even played a real song into it that did not work.
worthfreakingless
It would be a pretty amazing app if it was executed better/more accurate. It would be quite entertaining.
Actually, my music library is constantly growing thanks to Shazam. It only fails to identify live music (which sometimes it still gets even that right), old music, sometimes brand new music, and music by no name artists.
The only improvement I would like to see on Shazam is a different initial screen. Instead of starting the app by viewing all your recently tagged songs, I wish it would start with two large buttons: one for your tags and another to tag a song. At times, the small button to tag a song is hard to tap when you're in a rush to capture the last bits of a song on the radio or at the mall or something. I don't always start Shazam to look at my tags anyway....that only happens about <10% of the time.
And btw, I concur with the majority. Midomi is absolute crap. Only works less than half the time on popular songs.
The only improvement I would like to see on Shazam is a different initial screen. Instead of starting the app by viewing all your recently tagged songs, I wish it would start with two large buttons: one for your tags and another to tag a song. At times, the small button to tag a song is hard to tap when you're in a rush to capture the last bits of a song on the radio or at the mall or something. I don't always start Shazam to look at my tags anyway....that only happens about <10% of the time.
And btw, I concur with the majority. Midomi is absolute crap. Only works less than half the time on popular songs.
While I agree that Shazam is an incredible, "magic" app, I don't know about that "under heavy noise" thing. I was recently in a certain national chain steakhouse local store and two separate iPhones couldn't get Shazam to deal with music playing over their piped-in Sirius station, presumably because of background noise or maybe (comparatively) low volume level, or both. (Compared to my living room, I mean.) It wasn't that the songs weren't in the database, either, it was that the server couldn't "hear" what was playing.