[bigiain]

Are you sure?

From: https://www.theverge.com/circuitbreaker/2017/10/4/16408962/n...

"And this year’s Pixel will take advantage of the phone’s always-on microphones to listen for music (not just the phrase “OK Google”) and display what you’re listening to on the screen, even if it’s something on the radio."

Pretty sure it can't do that without at least transmitting audio fingerprints to the cloud passively...

Idea: a device that sniffs wifi and bluetooth MAC addresses and warns you when there's a Google Pixel2 in earshot...

(Note: That's the Pixel2, not the earbuds, but for the "magic" to work, the earbud wearers will have one of those in their pockets...)

[tkxxx7]

The feature you linked is also handled on-device.

[bigiain]

Really? I'd love to know how they do that? Can you _really_ cram a useable sized db of popular music fingerprints into something small enough to store locally on a phone these days?

[Shebanator]

I work for google but have no clue what underlying tech is being used here. But I have some familiarity with audio fingerprinting, so I thought I'd comment.

Using a not particularly efficient but reliable fingerprinting algorithm such as http://www.ismir2002.ismir.net/proceedings/02-FP04-2.pdf, you need 2.6kbits/sec for the fingerprint. Popular songs tend to be short, so lets say 3:30 per song on average. That works out to about 70K per song, or 70MB per 1000 songs. But compression and/or other sorts of encoding cleverness could massively reduce this number. Bottom line is that it wouldn't be hard to store thousands of fingerprints given a few hundred megabytes of storage (out of 64GB or more).

[pedalpete]

You don't need then entire song for a fingerprint, and the MusicBrainz (MetaBrainz) project has long been supported by Google.

I couldn't find anything which said how big their db is, but you could do some artist/song popularity smarts to figure out the tracks a user is most likely to listen to.

But to constantly be searching through fingerprints for every sound? I'd be interested to see what Googles solution to this may be.

[imaginenore]

There were 75K albums (not just songs) released in 2010 alone. The total number of songs is in hundreds of millions.

No way you can fit their fingerprints and the metadata into 64GB, compressed or not.

Shazam has 40M fingerprints in their DB, and they definitely don't have everything. Their software runs on beefy servers.

[pgeorgi]

The claim in the presentation was "the 10k most popular songs", updated weekly.

That won't satisfy the jazz nerd's needs, but they probably know their entire catalog by heart anyway.

[archagon]

I mean, would that satisfy a majority of needs, even? When I use Shazam, it's usually to identify a weird instrumental or otherwise obscure song that I've never heard before.

[aiiane]

Yes.

[hudbuddy]

I'm a bit late, but I watched the announcement for this feature and they explicitly said that nothing was sent to that cloud to determine what song it was.

I remember this stood out to me, because that was a large concern of mine when I first heard about the feature. I respected that they went out of their way to make this point.

She said that it is based on machine learning, and something like a database of "audio samples". it uses the samples to figure it out, I guess.

[kllrnohj]

> the earbud wearers will have one of those in their pockets

Use a recording app start recording the mic on your phone, put it in your pocket, and then talk at normal conversation volumes. It's not gonna pick up much, to say nothing of a person on the other side of the room.