Hacker News Reader

Comment by cannam | Hacker News Reader

I agree with you, and I work in music audio signal processing.

Shazam was founded in 1999-2000 ish and was live by 2002, longer ago than many people realise. The fingerprinting method they used was a novel one, and they successfully deployed it in a couple of years, with a database of two million tracks, in a system that could be used by dialling a number from any phone. It's an extraordinary and rare example of taking a research method and getting it to work at scale in the face of real constraints, in the process producing something that most people would never have imagined could be done.

View on HN

I was surprised that they didn't use a Jaccard distance when I looked into the paper, but I never really tested out why their more complicated method was required. The core insight (that note timing / beat is a fingerprint) is money, but the rest of the math seemed unnecessarily complicated.

Would really be interested if someone has tried other approaches and can comment.

THAT is the kind of startup I wish I could work for.

Me too. Are there any similar startup with functional prototypes or products? Something like Shazam, or Theranos if it were actually legitimate and had a working product.

ipfs?

SoundHound would fit the bill.

Echo Nest was like them, but they got gobbled up by Spotify several years ago.

Philips had similar technology that predates Shazam. That said, Shazam's algorithm scales better as they use "combinatorial hashing". This coupled with the fact they can up the hashes in a key-value table makes them better.

For anyone else still reading, this paper from Philips Research describes the method I think you're referring to. (It's cited in Avery Wang's 2003 Shazam paper.) http://www.ismir2002.ismir.net/proceedings/02-FP04-2.pdf

This is a really nice paper as well, but the Shazam method is more exciting because of its focus on the search part of the problem. Audio fingerprinting wasn't an entirely new field of course - the Philips paper also cites several prior publications.

The big limitation of all these methods is that they are robust to EQ and other degradations but not at all to variations in timing between frames. So they work well for identifying different instances of the same recording, but not at all for matching different performances, or for "query-by-humming".

Reply on news.ycombinator.com