[K0SM0S]

I've always been curious about Spotify's approach on this, because truth be told the service on good quality speakers does sound rather good.

I'm very pleased by everything I just read in this link — they really think of the user first, and favor music quality and the general experience. It's exactly how I did/do it myself, manually.

Awesome, really awesome to learn that the loudness war is finally coming to a close.

[strictfp]

I don't know how normalization is done, but I do know that Spotify are very professional when it comes to audio quality. They do double-blind testing with setups by audio engineers to evaluate end-to-end quality when evaluating changes.

[K0SM0S]

The double-blind scientific approach to audio testing ("ABX") has long been a major point of contention between "objectivists" and "subjectivists"[1]. This has been a godsend for the marketing marketing of some snake-oil companies in audiophile circles.

Really glad to know Spotify is on the side of science, here.

[1]: This blog has got to be the most interesting I ever read on whatever topic it touches, including the objective vs subjective debate: http://nwavguy.blogspot.com/2011/05/subjective-vs-objective-...

[sangnoir]

Do you mean double-blind testing benefits snake-oil companies? If so, may you expound on that? To my understanding, double-blind tests are meant to avoid un/conscious bias during testing.

[K0SM0S]

No, on the contrary! Sorry for my ambiguous writing. Indeed double-blind is the only way to go.

What I meant is that there is a sizeable chunk of the audiophile industry (a certain press, electronics brands, stores and even studios themselves) that conveniently avoids any and all scientific testing, and promotes typical snake-oil "features" and "specs". Quite sadly for their abused customers.

One of the worst trends in my opinion was faking technical format quality by using different masters on each — with the shitty master on CD/MP3 and the good one, more dynamic, on DVD/SACD/vinyl etc. The latter always sounded better simply because it wasn't the same source!

In such matters, I guess we can trust Spotify based on their claims, and anecdotally I tend to believe them.

[sramsay]

I assume they set a target of, say, -14 LUFS (Loudness Units Full Scale), take a reading of the integrated value for the whole track, and then turn the volume of the whole track down by whatever dB value will hit the target.

This avoids the problem of cruder methods, where you take, say, the peak dB as the "loudness" of the track. Doing it that way would take a fairly quiet song with one overly dramatic drum hit and make it so you couldn't hear the rest.

EDIT: I think they will also turn a track up if it's under the target, but I'm not absolutely sure. There would be no real reason not to.

[K0SM0S]

I think they mentioned ReplayGain with AAC and ogg formats, but it's possible they use some custom implementation. There's a good deal of information[1]; it is indeed not peak normalization:

> ReplayGain is different from peak normalization. Peak normalization merely ensures that the peak amplitude reaches a certain level. This does not ensure equal loudness. The ReplayGain technique measures the effective power of the waveform (i.e. the RMS power after applying an "equal loudness contour"), and then adjusts the amplitude of the waveform accordingly. The result is that Replay Gained waveforms are usually more uniformly amplified than peak-normalized waveforms.

Generally we indeed target both up or down in loudness (although 'up' is quite rare, typically high-end recordings of classical or jazz or otherwise very dynamic pieces).

Honestly, everything I read screams that Spotify really hired the right people to make these decisions.

Now I only have to rent about their handling (or rather, lackthereof) of metadata — how hard is it to display an actual first release date? but meh. Audio quality is great, no question about it. I just wish they'd stream FLAC in very-high quality + "quiet" environment settings (the "pure hi-fi" experience so to speak), but that'll come in due time I suppose.

[1]: https://wiki.hydrogenaud.io/index.php?title=ReplayGain