|
|
scott wrote:
> Actually I have a phone book entry on my phone, which you can call, hold
> your phone up to some random speaker (be it in your car, on the TV, etc)
> for 10 seconds, then a few seconds later it texts you back the artist
> and song name. It's pretty neat and has worked every time I've tried it.
I just finished software doing that, except it would also recognise what
radio station you're listening to, whether it was an advertisement and from
whom, or whether it was a talk show and who was talking.
It would also do TV, and give you (for example) links to the MTV episode
information, or the buy-this link if you listened to the Home Shopping Network.
(I'm hoping the bosses find some more money, but it's looking grim still.)
The audio identification is all based around taking a short sample and
finding it amongst a set of larger samples. It's pretty freaky, because the
algorithm even finds (for example) people covering songs. We had an
American Idol caller come in and the system said "I don't know who is
singing this, but it's Yesterday by the Beatles." :-) Pretty cool.
So the Fraunhoffer stuff is recognising music based on frequency analysis,
beat detection, and a couple other parameters that are of course secret. The
signatures it generates are about 300 bytes/second of audio, with even less
than that needing to actually be in memory at any given time for a
preliminary match. But given that it's a linear size increase, I'm guessing
it's something not unlike string matching, except with fuzzy audio
characteristics. It's usually either a very good match, or one of these
"Could be Celine Dijon, could be Mozart, could be Ozzy Osborne" kind of things.
> Maybe it works in the frequency domain, so takes the fourier transform
> of the sample, then uses some fuzzy matching algorithm to see what it
> matches up with?
Basically that sort of thing, yes.
>> (But then, so should identifying a CD by it's serial number, and
>> apparently that is a "solved problem".)
>
> What do you mean? Isn't the whole idea of a *serial* number that you can
> identify which one it is?
CDDB works by looking at track lengths and positions. I.e., the information
in the table of contents that says what sector each song starts and stops
on. Audio CDs for some reason don't have serial numbers, UPC codes, lyrics,
or cover art encoded on them, which seems like rather a loss to me. (Even
stranger to me is that MP3 doesn't define that sort of thing either.)
--
Darren New, San Diego CA, USA (PST)
The NFL should go international. I'd pay to
see the Detroit Lions vs the Roman Catholics.
Post a reply to this message
|
|