Does normalizing require upconverting to preserve quality?

[Robert J. H.]

When you amplify in 16-bit in the first place, The result will be the same as with multiplying in 32-bit float and re-converting. The rounding-off error is the same. But with the dithering, you can preserve some of the accuracy of the 32-bit precision.
One might think "Alright, I'm going to apply this trick in each calculation" - even in 16-bit mode.
Unfortunately, this works only once.
We use 1 bit for the dithering. This could amplify the noise in the next process by 2 (when the random noise is exact the same the second time) but in General by 1.41.
An example: You can create 5 tracks of pink noise and Amplitude 0.1.
If you only duplicate the tracks, the Amplitude will be 0.5 when mixed down into one track. I've tried it with 5 different tracks and the final Amplitude was 0.27.
To summarize it:
- It doesn't matter if you amplify a 16-bit wave form with 16 bit precision or if you convert it to 32 to do the calculation and then go back.
- example: Multiplikation with integer or anything below 16-bit that does not introduce new bits.
- exeption to this rule would be if you multiply 16-bit with 16-bit or 16-bit (as 32) with 32-bit).
- 16-bit is already high Quality, the aliasing is therefore only perceivable when the passage gets quieter and fewer bits can be used to express the Dynamics (compare the fade-out example from the dithering guide).
- It is therefore important to Keep the gain Close to 0 db while leaving some dB headroom for dithering (and resampling, if needed).
- Sounds that are made with 32-bit precision Sound better when dithering is applied just before converting to 16-bit.
- triangle shaping is the quietest one (measureble peaks, not perceived loudness) and is less fatal than other types, when applied multiple times (bad Thing in General).
- Psychoaccoustic shapes are best suited for final dithering.

[steve]

But what if I edit in 24 bit?

24 bit is more accurate than 16 bit but less accurate for processing than 32 bit float.
Quantize distortion is less when processing in 24 bit than processing in 16 bit, but not as good as processing in 32 bit float.
The amount of dither required (and so applied) when changing from 32 bit float to 24 bit is much less than when changing from 32 bit float to 16 bit, but if you then convert from 24 bit to 16 bit, you still need to apply dither if you are to avoid harmonic distortion due to quantize errors as rounding is required whenever a 24 bit sample cannot be converted exactly to 16 bit.

If the "multiplication" was done in 16 bit in the first place, there would be no problem to begin with.

Yes, but that is just one special case for one type of process. What if you are amplifying by a negative amount (to make the audio quieter), or filtering, or fading or noise reduction, or DC offset removal....? In these cases it is highly unlikely that each processed sample will be an exact 16 bit value, so there will be quantize errors.

A better way to avoid the need for dither, is to keep everything in 32 bit float format (or higher). If you don't convert down to a lower bit format, there is no need to dither. The problem there is that few programs (other than audio editing programs) support 32 bit float format (also, the file size is twice as big as a 16 bit file).

I normally use shaped dither.

[KlarkKentThe3rd]

I still don't understand why someone couldn't just invent a way of doing amplification (positive and negative) in 16 bit without producing errors.

Like:
1 3 5 7 9 11

Aplify by 200%

2 6 10 14 18 22

Why is that impossible again?

[steve]

Sure that will work for amplifying by 200%, 400%, 800%...

Your original question was about normalizing, so what happens when normalizing requires 3.287341 dB amplification?

[KlarkKentThe3rd]

Sure that will work for amplifying by 200%, 400%, 800%...

Your original question was about normalizing, so what happens when normalizing requires 3.287341 dB amplification?

I had time to think about that. I understand now: you need fractions for more precise actions.

And I realized I don't need normalizing anyway; reducing bass was the solution all along.

And as long as we are on the topic of dithering, can it only be applied to the area I edit? I got a bunch of tracks that are in a desperate need for a proper fade out, but I am really concerned with leaving them as untouched as I can. Can dither be applied to the selected section only, and not again on export? That is the last thing I want to learn about dithering in Audacity.

[Gale Andrews]

A better way to avoid the need for dither, is to keep everything in 32 bit float format (or higher). If you don't convert down to a lower bit format, there is no need to dither. The problem there is that few programs (other than audio editing programs) support 32 bit float format (also, the file size is twice as big as a 16 bit file).

Current versions of Windows Media Player (Windows 7 and 8) and iTunes (on Mac and Windows) should support 32-bit float WAV now ... if you have the disk space.


Gale

[Robert J. H.]

Sure that will work for amplifying by 200%, 400%, 800%...

Your original question was about normalizing, so what happens when normalizing requires 3.287341 dB amplification?

I had time to think about that. I understand now: you need fractions for more precise actions.

And I realized I don't need normalizing anyway; reducing bass was the solution all along.

And as long as we are on the topic of dithering, can it only be applied to the area I edit? I got a bunch of tracks that are in a desperate need for a proper fade out, but I am really concerned with leaving them as untouched as I can. Can dither be applied to the selected section only, and not again on export? That is the last thing I want to learn about dithering in Audacity.

There's a bunch of dithering Plug-Ins out there.
Since Plug-Ins can only apply their effect to the selected Audio, the rest of your track will be untouched.
However, I don't know how distinguishable the boundaries are in such a case. Perhaps the noise Level is so deep that it won't be audible.
Maybe Steve and I could write a separate plug-in if he's not too much snowed under already.

[steve]

Since Plug-Ins can only apply their effect to the selected Audio, the rest of your track will be untouched.
However, I don't know how distinguishable the boundaries are in such a case.

It's not likely to be very noticeable, but given that we are quibbling about details that are barely noticeable, that would be my concern. Ensuring that transitions between using dither and not using dither only occur when there is a reasonably high signal level should mask the transition.

And as long as we are on the topic of dithering, can it only be applied to the area I edit?

We need to make the distinction between "editing" and "processing".
A simple edit (Cut / Paste / Delete / and so on) do not cause sample values to change.
"Processing", such as amplifying, fading, reducing the bass or treble, increasing the bass or treble, or any other "effect", do change the sample values.

What you would need to do:
1) In Preferences, set the Default Sample Format to 16 bit and enable dither for "High quality conversion".
2) Import the 16 bit file that you want to work with.
3) Perform edits and processes as required. "Processed" audio will have dither applied automatically.
4) Disable dither before exporting.

Note that if you apply two or more processes to the same bit of audio (such as reducing the bass and applying a fade), the result will be worse than if you had just left Audacity with its default settings.

[KlarkKentThe3rd]

Thanks everyone, for your help.

I took a bass-heavy track, reduced the low frequencies by 2 db, and applied triangular dither during the process. The end result sounds perfect.

[doubleO7]

Hello everybody,

here is a little video that might clarify a few things about the behaviour of digital audio signals including dither.

Cheers Christian