Mastering technical stuff

[Atlantis]

1 - Could you explain the difference between Linear Phase Equalization is a opposed to non-Linear phase EQ?

Basically, any analogue or digital EQ introduces phase shifts to different frequencies, which can smudge transients over time and further colour the sound. This may be desirable in some cases, but when performing critical EQ correction, it is better to use a linear-phase EQ which doesn't introduce any phase shifts, keeping the signal pure (apart from lowering or boosting frequencies).

2 - I'd be interested to know the specific model of your monitors.

Dynaudio BM5A. I've had them for years and am so used to their sound.

[Atlantis]

Or how about simply using FLAC? Lossless audio codecs compress audio signals better than lossless general-purpose compression algorithms. Apart from that, the 7z format should be preferred over RAR at any time, since it's public domain and has a very similar compression ratio.

I totally agree. But most people are unaware of the format, at least I think so. I will include it in my set of steps though.

EDIT: FLAC doesn't support floating point, does it? That's why I liked Monkey's Audio, but it's just not as popular.

[Saga Musix]

Yeah, FLAC is fixed point only, but since there is a variety of open lossless formats to choose from, everyone should be able to find something that suits them best.
BTW: As long as there is no clipping, it doesn't really matter if you use floating point or not. The only advantage of floating point over fixed point is during production, as you can avoid clipping when routing audio through various plugins/etc.... However there is really no difference when dealing with normalized files. Afterall, a 32-bit floating point PCM signal has the same SQNR ratio as a 24-bit fixed point PCM signal (because the mantissa of 32-bit floating point numbers happens to be 24 bits wide).

[Atlantis]

Yeah, FLAC is fixed point only, but since there is a variety of open lossless formats to choose from, everyone should be able to find something that suits them best.
BTW: As long as there is no clipping, it doesn't really matter if you use floating point or not. The only advantage of floating point over fixed point is during production, as you can avoid clipping when routing audio through various plugins/etc.... However there is really no difference when dealing with normalized files. Afterall, a 32-bit floating point PCM signal has the same SQNR ratio as a 24-bit fixed point PCM signal (because the mantissa of 32-bit floating point numbers happens to be 24 bits wide).

That's true. But what if a track clips internally and you still render in fixed point? Mind you, I never understood the notion of clipping the sound altogether, since I believe certain plug-ins work in fixed point anyway, and if you keep the level low enough, clipping just doesn't occur. And what about low-level signals - doesn't floating point provide greater resolution there, too? That you say it has the same SQNR makes me think otherwise.

[KrazyKatz]

Basically, any analogue or digital EQ introduces phase shifts to different frequencies, which can smudge transients over time and further colour the sound. This may be desirable in some cases, but when performing critical EQ correction, it is better to use a linear-phase EQ which doesn't introduce any phase shifts, keeping the signal pure (apart from lowering or boosting frequencies).

I wasn't aware that it's actually possible to EQ without introducing phase shifts, hence my asking. It almost doesn't make sense. I'll have to look into it more.

I'd say this thread somehow needs to also be in the Technical Documents. A lot of useful info here.

[Saga Musix]

That's true. But what if a track clips internally and you still render in fixed point?

If we assume that you get your jobs from OpenMPT users as you're asking on this forum, it won't make a difference because OpenMPT's internal mixing resolution is fixed at 32-bit integer (which has a higher SQNR than 32bit floating point, but a lower dynamic range because floating point won't clip that easily)

I believe certain plug-ins work in fixed point anyway

If we're talking about VST plugins here, that would be very bad practice, since the VST interface is defined to work with floating point data - so I assume most plugs work with floats internally as well.

And what about low-level signals - doesn't floating point provide greater resolution there, too? That you say it has the same SQNR makes me think otherwise.

Since the SQNR is the same as with 24-bit integer, you get the same noise level at low levels.

[Atlantis]

If we assume that you get your jobs from OpenMPT users as you're asking on this forum, it won't make a difference because OpenMPT's internal mixing resolution is fixed at 32-bit integer (which has a higher SQNR than 32bit floating point, but a lower dynamic range because floating point won't clip that easily)

I understand that, yes, but most mixes come from the popular programs such FL Studio and Logic.

If we're talking about VST plugins here, that would be very bad practice, since the VST interface is defined to work with floating point data - so I assume most plugs work with floats internally as well.

I've always wondered that, because Waves claims that some of their plug-ins work using 48-bit double precision and truncate to 24 bit (you can choose wether to use dither or not).

Since the SQNR is the same as with 24-bit integer, you get the same noise level at low levels.

Thanks for clearing that up.

[Saga Musix]

I've always wondered that, because Waves claims that some of their plug-ins work using 48-bit double precision and truncate to 24 bit (you can choose wether to use dither or not).

They might work with 64-bit floating point data internally then... Though the mantissa is 53 bits wide for 64-bit floating point, not 48 bits... so I don't know what they are exactly doing.

[Atlantis]

They might work with 64-bit floating point data internally then... Though the mantissa is 53 bits wide for 64-bit floating point, not 48 bits... so I don't know what they are exactly doing.

So the VST spec is designed to work in 32 bit floating point? I've always wondered that because I always upsample and work in 64 bit floating point, which some plug-ins work at. That those Waves plug-ins work in 48-bit double precision could work too then, provided they truncate or dither it back to 24 bit (which they do). So, basically, I can omit having to render in 32 bit floating point and just state 24 bit? What if the track clips internally though?

[Saga Musix]

So the VST spec is designed to work in 32 bit floating point?

VSTs and hosts are "talking" in 32-bit or 64-bit (only possible since VST 2.4 and only if both the host* and the VST support it) floating point precision.
If there is a difference between 32-bit floating point and 24-bit fixed point is largely dependent on the host and the plugins used; in OpenMPT I'd for example say that rendering in 24-bit quality might be better if there is no clipping happening (because it internally renders in 32-bit fixed point precision, so all what is needed for rendering is to dither or cut off the least signifcant 8 bits). Basically you should avoid internal clipping to happen anyway, so I dunno if there would be a different in OpenMPT, in other hosts there might be of course.
I'd say there are good arguments for both 24-bit fixed point and 32-bit floating point, and I wouldn't decide on a winner at all.

*OpenMPT does not support 64-bit floating point communication.

[Atlantis]

So the VST spec is designed to work in 32 bit floating point?

VSTs and hosts are "talking" in 32-bit or 64-bit (only possible since VST 2.4 and only if both the host* and the VST support it) floating point precision.
If there is a difference between 32-bit floating point and 24-bit fixed point is largely dependent on the host and the plugins used; in OpenMPT I'd for example say that rendering in 24-bit quality might be better if there is no clipping happening (because it internally renders in 32-bit fixed point precision, so all what is needed for rendering is to dither or cut off the least signifcant 8 bits). Basically you should avoid internal clipping to happen anyway, so I dunno if there would be a different in OpenMPT, in other hosts there might be of course.
I'd say there are good arguments for both 24-bit fixed point and 32-bit floating point, and I wouldn't decide on a winner at all.

*OpenMPT does not support 64-bit floating point communication.

That's exactly how I think - clipping just should never occur! Thanks for your explanation though. There was an article discussing the benefits of floating point over fixed point (or the other way around), but I don't think I have it bookmarked anymore. In reality though, it's hard enough to distinguish 24 bit from 16 bit, so why bother with anything greater?

What I'm doing now, is resample my mixes to 96,000 Hz, 32 bit float using r8brain PRO, upsample to 64 bit floating point, then process them in Sound Forge, resample them again using r8brain PRO (which truncates to 32 bit float), and dither to 16 bit after that. I can't see how I could improve that process any further.

[LPChip]

I've split the topic in two. Couldn't find a good title, so Atlantis or any other staffmember, feel free to edit it to a better title.

I was first thinking of Mastering mumbojumbo, but I didn't think that was better as what I have now.