Like I said in another thread, it's what the client has asked for. I can advise on good sound/best practice, but at the end of the day it's their baby, not mine.

A more constructive note: We need a distribution format that doesn't have a clipping limit. Lets give the loudness warriers a free run, and remove the barrier they have been banging their heads against for decades. This will practically force the issue of normalizing tools in end-user devices. It is a way to make it worse in order for it to get better. As long as we have the ubiquitous clipping barrier at 0dBFS, the pain is not great enough to include loudness control in end user devices, but if the distribution format doesn't clip anymore, it soon will.

Isn't 32-bit float the new up-and-coming audiophile format anyway?

Isn't 32-bit float the new up-and-coming audiophile format anyway?

Is it? Perhaps I'm a bit behind. Does the Pono player understand it?

Pono to play dsd

Quote from: Wombat on 2015-01-24 17:16:43

Pono to play dsd

That's not the same, is it?

A more constructive note: We need a distribution format that doesn't have a clipping limit. Lets give the loudness warriers a free run, and remove the barrier they have been banging their heads against for decades. This will practically force the issue of normalizing tools in end-user devices. It is a way to make it worse in order for it to get better. As long as we have the ubiquitous clipping barrier at 0dBFS, the pain is not great enough to include loudness control in end user devices, but if the distribution format doesn't clip anymore, it soon will.

This doesn't work. Normalized float would still be +/-1.0 = 0 dBFS, and nobody in his right mind would hard clip to that. So any gain that the mastering engineer adds above 0 dBFS would be pointless, since it would be reverted at the consumer.

What we need is the industry to agree on a certain loudness target, for example -18 LUFS (like ReplayGain uses). And before that can happen we need to enable ReplayGain in all players..

• The clipping level differs between different pieces of gear. Potentially by a significant amount.
• The clipping is not as hard as with digital audio, and hence produces less energy in the higher harmonics.
• The ultrasonic harmonics don't alias down into the audible band.

There aren't any widely adopted floating point ADCs or DACs around, that would kind of scupper things, no?

Quote from: darkflame23 on 2015-01-24 19:51:12

There aren't any widely adopted floating point ADCs or DACs around, that would kind of scupper things, no?

I don't see why. You don't necessarily feed the signal from a stream directly into a DAC, do you?

Well, I guess your customers are individual artists.
But it is understandable that even they want a master that is as loud as similar recordings, which can mean loads of compression. It's a vicious circle.


I said this many times before, the solution to this is ReplayGain enabled by default in the major audio players (soft- and hardware). Punish loud tracks.

As soon as a majority agrees on such a target, there is a heightened incentive for someone to violate it. Louder than everything else still rules, and will dominate every agreement you could possibly negotiate.

Removing the 0dBFS barrier will allow mastering engineers to keep the peaks without sacrificing loudness, and of course they will exceed +/-1.0 in the process. With floating point, there's no harm in that.

I'm sure that many engineers saw it as a feature, not a problem, that digital audio has such a well defined level limit.

I would not buy a player that uses ReplayGain.

You do if you actually want to listen to it.

Didn't want to start an argument. If it's gonna end up as fixed point when you hear it, I was just wondering what the advantage of 32bit float as a distribution medium would be. Genuinely interested to know.

I thought I had explained just that. I seem to have failed. The removal of the 0dBFS limit, remember? Once the signal is under the consumer's control, the consumer can adjust the level to his own liking (by using manual volume controls or by using automatic level equalization software) and then convert to fixed point safely.

To put the salient point in other words: The difference is that the clipping isn't done on the producer end, it is done (or better: avoided) on the consumer end. To make that possible requires a different distribution format.

if it robs the content producers of their main reason to compress their product to death.

You don't need floating point for that.
Getting the point of lower loudness across should be hard enough. Adding a new sample format on top of that will make the acceptance virtually impossible.

No, because ReplayGain in the device would automatically punish everything louder than the target.

At this point the target is not agreed upon to reduce loudness per se but to ensure enough headroom for dynamics. Sure, you can squish the waveform and compress the sh*t out of it as an artistic choice, but at the consumer an e.g. -8 LUFS track would then be punished by -10 dB gain (to reach the -18 LUFS target). You'd get squished sound that would be just as loud as a much more dynamic track.

This doesn't work. You cannot preserve the peaks and increase loudness. At the consumer the audio has to be normalized to +/-1.0 so if the engineer keeps the peaks above 1.0 it will cause low volume during playback due to normalization.

Unless you have something like ReplayGain, the engineers would still simply squish the waveform and limit to +/-1.0.

Instead of fuzzy limits you get clean sound all the way up to the limit.

The problem comes with the abuse, which could have been prevented with a common loudness target in the past. With that we wouldn't even need ReplayGain nowadays.

Dynamic range compression isn't hard clipping and hard clipping should be avoided, so you will at least need normalization. The result is identical to our current situation, except that you need extra processing power.

You failed to explain this: why the producers will not compress audio?

With floating point, they can go over 1.0, and have more loudness if they want.

The consumer also wins: He gets the unsquashed product with good sound quality (if the producer so wishes), albeit at a possibly too high playback level.

Yes, but you'd need to make ReplayGain ubiquitous before, and there currently isn't enough incentive for that, because the 0dBFS limit achieves a kind of level matching already, albeit at the cost of a lot of distortion.

You don't understand. With floating point, the content producers don't have to be convinced of lower loudness. With floating point, they can go over 1.0, and have more loudness if they want. For them, it is the removal of a previous restriction. Plus they can have what they pretend to have wanted all along: Good sound and more loudness at the same time. The orgy without the hangover. You see: All pluses, no minus!

The consumer also wins: He gets the unsquashed product with good sound quality (if the producer so wishes), albeit at a possibly too high playback level. That's easy to correct: He turns down the level. If he's annoyed by having to do that, he'll use ReplayGain or something similar. That'll help to make automatic loudness control systems ubiquitous. I hope it is not going to be ReplayGain only. A competition for the best system is good here, too. That's a competition the market can conduct without any cooperation from the content producers.

Yes, but you'd need to make ReplayGain ubiquitous before, and there currently isn't enough incentive for that, because the 0dBFS limit achieves a kind of level matching already, albeit at the cost of a lot of distortion. People will only convert to ReplayGain in masses once the levels of the different titles they listen to vary to such a degree, that it becomes a pain in the bottom part of the back.

This means that you have to agree an amount of headroom that is valid for everyone. That is going to be a tough one to agree on. With floating point, that's entirely unnecessary. The distribution format has almost infinite headroom. That means the distribution format disappears as a factor from the equation. Isn't that greatly superior?

No, the consumer simply turns the volume up or down according to his preference, either manually or automatically. That's an operation that still uses floating point, so nothing is lost. Only then the conversion to fixed point, and therefore the clipping to +/- 1.0, takes place. It depends on the consumer's playback system where this clipping point is relative to the overall loudness. The content producer has nothing to do with it anymore.

I don't see why they would. They can't anticipate what volume the consumer would choose. They gain nothing from clipping at 1.0 at their end.

Yes, and you get everybody banging their heads against that limit in the quest for the loudest record ever. At the expense of clean sound.

We had that target more or less informally at the beginning of the CD era. The loudness war has done away with that. The lesson is that no scheme will work that relies on the content producers sticking to the rules. They won't. It is not in their interest. So yes, the problem comes with the abuse. We all know. But that doesn't make it go away.

Dynamic range compression isn't hard clipping and hard clipping should be avoided, so you will at least need normalization. The result is identical to our current situation, except that you need extra processing power.

It is not identical, because it doesn't happen at the same point in the distribution chain, and therefore not controlled by the same people and interests. Don't you see the difference between normalization done by the mastering engineer, and normalization done by the consumer?

He's proposing a paradigm shift where float is converted to integer after distribution and after mandarory loudness equalization.