Best Dither, Sample-Rate and Bit-Depth for Masters?

[paulsmusic]

I've recently started using Fabfilter Pro-L2, and the manual states that dithering isn't necessary when creating MP3 masters.

I've always created all my master files at 16-bit / 44.1Khz, with dither applied, so that they would be CD compatible. In Cubase, my mastering projects are normally 24-bit at 48Khz (although I've recently switched to 64-bit floating), and I do the final bit-rate / sample-rate conversions when I bounce the master.

I'm thinking that for digital masters (intended for online distribution / streaming), it would be better using 24-bit / 48Khz (not dithered?), and for CD masters, 16-bit / 44.1Khz dithered.

I'd be grateful for advice on recommended final bit-rates / sample rates, and whether or not to apply dither.

[Hugh Robjohns]

Assuming your aim is to preserve audio quality, if you reduce the wordlength you need to dither. Thats the rule. No exceptions.

There is a trend towards mastering at 96kHz, 24bits as a 'high-res' source.... but 48 or 44.1kHz are both perfectly acceptable for most material and there's negligible quality difference between them. Pick whichever bests suits your workflow and preferred streaming service.

If you still produce a lot of CDs then it would be easier to run projects at 44.1kHz (24bit) just to avoid the need for sample rate conversion (you'd still need to dither down to 16bit of course).

[RichardT]

You need to dither whenever you reduce the word length (e.g. from 32 bits to 24 bits). Changes to the sampling rate are not relevant to whether you need to dither.

In Cubase, there are two main settings to consider

- 1. the word length of recorded audio files, as stored in the pool.
- 2. the number of bits used by the audio engine (32 or 64)

No matter what you choose for the word length of the recorded audio in 1., Cubase will process audio internally at the number of bits selected in 2. So internally, the audio is always at 32 or 64 bit.

That means you always need to dither on audio export. Cubase does not dither automatically, unlike many DAWs.

For streaming services, either 44 or 48 is fine, but you should dither on audio export. I use 24 bits for streaming masters.

In Pro-L, I use the optimised dither, which is a noise-shaping dither that moves most of the noise to higher frequencies. This is assuming that what you export from Cubase is not processed in another software package before it’s distributed.

[paulsmusic]

You need to dither whenever you reduce the word length (e.g. from 32 bits to 24 bits). Changes to the sampling rate are not relevant to whether you need to dither.

In Cubase, there are two main settings to consider

- 1. the word length of recorded audio files, as stored in the pool.
- 2. the number of bits used by the audio engine (32 or 64)

No matter what you choose for the word length of the recorded audio in 1., Cubase will process audio internally at the number of bits selected in 2. So internally, the audio is always at 32 or 64 bit.

That means you always need to dither on audio export. Cubase does not dither automatically, unlike many DAWs.

That's so helpful, thank you! I normally have a recording / mixing project, with the recording bit-depth set at 24-bit. I export this at 24-bit to use in a mastering project for final file conversion...so, are you saying that I should be dithering the recording project file when I export it, as Cubase is using 32-bit/64-bit internally? Wouldn't this mean that I would then have to dither the file again when I do my final master?

[paulsmusic]

Assuming your aim is to preserve audio quality, if you reduce the wordlength you need to dither. Thats the rule. No exceptions.

There is a trend towards mastering at 96kHz, 24bits as a 'high-res' source.... but 48 or 44.1kHz are both perfectly acceptable for most material and there's negligible quality difference between them. Pick whichever bests suits your workflow and preferred streaming service.

If you still produce a lot of CDs then it would be easier to run projects at 44.1kHz (24bit) just to avoid the need for sample rate conversion (you'd still need to dither down to 16bit of course).

Thanks Hugh, makes sense to me!

[Hugh Robjohns]

Wouldn't this mean that I would then have to dither the file again when I do my final master?

You dither every time the word length is reduced.

In effect, dither re-encodes the low level audio information into the new lowest bits that would otherwise be lost when the wordlength is reduced.

So yes, you dither when converting from 32 or 64bit floating point to 24 bit (fixed point), and again if converting from 24 bit to 16 bit.

[RichardT]

You need to dither whenever you reduce the word length (e.g. from 32 bits to 24 bits). Changes to the sampling rate are not relevant to whether you need to dither.

In Cubase, there are two main settings to consider

- 1. the word length of recorded audio files, as stored in the pool.
- 2. the number of bits used by the audio engine (32 or 64)

No matter what you choose for the word length of the recorded audio in 1., Cubase will process audio internally at the number of bits selected in 2. So internally, the audio is always at 32 or 64 bit.

That means you always need to dither on audio export. Cubase does not dither automatically, unlike many DAWs.

That's so helpful, thank you! I normally have a recording / mixing project, with the recording bit-depth set at 24-bit. I export this at 24-bit to use in a mastering project for final file conversion...so, are you saying that I should be dithering the recording project file when I export it, as Cubase is using 32-bit/64-bit internally? Wouldn't this mean that I would then have to dither the file again when I do my final master?

What Hugh says is exactly right.

To answer your specific questions, if you are recording at 24 bit and exporting at 24 bit, yes, you do need to dither, for exactly the reason you gave.

If you are mastering in a separate project, then the audio export from the mixing project should not use optimised dither but TPDF dither or an equivalent (assuming you’re not limiting at this stage, use the Lin dither plugin at 24 bits with noise shaping off).

When you export from the mastering project, you need to dither again, this time using optimised dither in Pro-L

Dithering increases the noise by 3dB each time, but with the enormous dynamic range of 24 bit, that’s not a problem.

[Mike Stranks]

The question was framed in the context of MP3s, before moving on to wider contexts.

I've read elsewhere that if creating MP3s from WAVs then you shouldn't apply dither... and that seems to be borne-out by iZotope's export function in RX. If exporting to MP3, you don't get the option to apply dither.

Just putting that out there as MP3s seem to be a different case to WAVs.

[paulsmusic]

So yes, you dither when converting from 32 or 64bit floating point to 24 bit (fixed point), and again if converting from 24 bit to 16 bit.

Thanks for clarifying Hugh! I'd read / seen somewhere that you weren't meant to dither twice...so much information on the internet, some good and some bad I guess!

[muzines]

In effect, dither re-encodes the low level audio information into the new lowest bits that would otherwise be lost when the wordlength is reduced.

This is a really helpful sentence to sum up what's going on here - thanks!

Some questions I'm not sure about:

I'd always dither when exporting at 24-bits or less, but let's say the DAW is using a 32-bit float engine, and we export 32-bit files. I'd expect not to need to dither in that case, as we are not reducing the wordlength. Ok.

If the DAW is using a 64-bit float engine (for example, Logic has the choice between 32 & 64), and we're exporting to a 32-bit file - we should probably dither, right, as we're supposedly reducing the wordlength? (Albeit with the noise floor extrememly low) But in Logic, dither algorithms are not available when choosing to bounce as 32-bit float (the only 32-bit bounce option).

Also - when we're just using the DAW while working on projects, with it's 64-bit floating point engine - the stereo output is going to the audio interface, generally at 24-bits fixed point by the time the audio leaves the DAW and heads to the interface - but DAWs aren't generally dithering their regular output as far as I'm aware, and thus more or less everyone would always be monitoring undithered wordlength reduction while working on their projects.

I wonder whether the DAW is doing no dithering, or perhaps some naive light-CPU dithering in this case? How can we find out? Does it matter?

My mind is somewhat dithering about all of this... help me, Obi-Hugh, you're my only hope!

[paulsmusic]

What Hugh says is exactly right.

To answer your specific questions, if you are recording at 24 bit and exporting at 24 bit, yes, you do need to dither, for exactly the reason you gave.

If you are mastering in a separate project, then the audio export from the mixing project should not use optimised dither but TPDF dither or an equivalent (assuming you’re not limiting at this stage, use the Lin dither plugin at 24 bits with noise shaping off).

When you export from the mastering project, you need to dither again, this time using optimised dither in Pro-L

Dithering increases the noise by 3dB each time, but with the enormous dynamic range of 24 bit, that’s not a problem.

Thanks Richard! These are the Noise Shaping Options in L-2:

The Noise Shaping setting lets you choose between various noise shaping algorithms:

The Basic setting lowers the overall noise floor a few dB, at the cost of increasing noise
levels for frequencies above 6 kHz.

With the Optimized setting, the effect is more extreme; you'll get an even lower overall
noise level, but noise frequencies above 10 kHz are boosted more extremely.

Weighted noise shaping will transform the noise spectrum according to the ear's sensitivity
to certain frequencies at low listening levels. Theoretically, this results in the lowest audible
noise. This noise shaping setting is designed to be used at 44.1 kHz. It still works at other
sample rates, but the frequency spectrum of the resulting noise isn't optimal anymore.

[Hugh Robjohns]

Just putting that out there as MP3s seem to be a different case to WAVs.

The audio file going into an mp3 encoder should be dithered at whatever its word length is. If it wasn't it would contain low-level distortion artefacts...

However, the mp3 encoding process uses different wordlengths in different frequency bands in a programme-dependent dynamic way, so dither really has no meaning

[Hugh Robjohns]

I'd read / seen somewhere that you weren't meant to dither twice...

You shouldn't dither twice at the same word length — the reason being that it would only raise the noise floor without doing anything useful.

But if you're changing the word length, you must dither.

A possibly confusing situation is if you import into a DAW a 24 bit file, say, and perform some function on it like a level change, EQ etc, and then export at 24 bits again. The word length is the same as the source, so no need to dither?

Except there is a need because the moment the file was imported it was converted to the DAW's internal format of 32 or 64 bit floating point. So exporting at 24 bit is a reduced wordlength and does need dithering.

[Hugh Robjohns]

I'd always dither when exporting at 24-bits or less, but let's say the DAW is using a 32-bit float engine, and we export 32-bit files. I'd expect not to need to dither in that case, as we are not reducing the wordlength. Ok.

I'd agree.

If the DAW is using a 64-bit float engine (for example, Logic has the choice between 32 & 64), and we're exporting to a 32-bit file - we should probably dither, right, as we're supposedly reducing the wordlength?

Yes... although the digital noise floor will be so far below the signals noise floor that it will self-dither anyway, so it probably doesn't matter in practice.

Also - when we're just using the DAW while working on projects, with it's 64-bit floating point engine - the stereo output is going to the audio interface, generally at 24-bits fixed point by the time the audio leaves the DAW and heads to the interface - but DAWs aren't generally dithering their regular output as far as I'm aware, and thus more or less everyone would always be monitoring undithered wordlength reduction while working on their projects.

I don't know about Cubase, but the SADiE DAWs I use most does dither the 24 bit signals going to the converters. I'm not sure about others...

[M Andersen]

Analog tape hiss dither can be recommended ; )

[RichardT]

Analog tape hiss dither can be recommended ; )

True - but the dynamic range of tape is equivalent to a digital signal of word length 13 bits or so!

[James Perrett]

I'd also suggest checking that your DAW is dithering properly. Dither is one of those processes that programmers often seem to overlook and I've had to report dither related bugs in the past.

While many places now accept high resolution files, not all do, so I still supply mainly 16 bit 44.1kHz files so that I know that I'm in control of dithering and any necessary sample rate conversion.

[paulsmusic]

While many places now accept high resolution files, not all do, so I still supply mainly 16 bit 44.1kHz files so that I know that I'm in control of dithering and any necessary sample rate conversion.

Thanks James...yes, I was wondering what happens if you supply a 24-bit master and it gets converted to 16-bit by a streaming service. Would they apply dither? I found the article below on Pro Audio Files which lists the formats used by different streaming services, and suggests mainly supplying 24-bit files, apart from CDBaby, who are currently still only accepting 16-bit...

https://theproaudiofiles.com/audio-mast ... EAMBk3LvQs