• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

Does DSD sound better than PCM?

Status
Not open for further replies.

DonH56

Master Contributor
Technical Expert
Forum Donor
Joined
Mar 15, 2016
Messages
7,835
Likes
16,497
Location
Monument, CO
My Earthworks M30 is specified flat from 3 Hz to 30 kHz. The higher model, the M50, is spec'd from 3 Hz to 50 kHz. There are other specialty mics from other companies with significantly higher frequency range (and prices to match).

https://earthworksaudio.com/microphones/m-series/
 

Miska

Addicted to Fun and Learning
Audio Company
Joined
Feb 20, 2019
Messages
615
Likes
448
No, the error processes are not the same thing as noise and distortion, which are outcomes.

Are you suggesting that the level of noise and distortion (relative to the signal) introduced by 32-bit floating are comparable in level to the noise and distortion (relative to the signal) introduced by the best analogue consoles and units?

Error -> distortion & noise. Yes, depending on what you do. That's why I use 64/80-bit floating point. Even 64-bit floating point is not always enough.

It is not hard to fuck up digital processing. By just doing something in digital doesn't mean it is automatically better, because digital has it's own problem domains. Poorly done digital can be easily worse than well done analog.

If you've ever used the "old school" computer music composition systems called "trackers", you know how much worse 16-track MOD sounded than 4-track one.

None of the DAW's / desks I've seen give out any estimate of remaining calculation precision after chaining certain number of plugins/settings combination. And loss of precision (distortion) sounds worse than increase in analog noise floor (noise).
 

Miska

Addicted to Fun and Learning
Audio Company
Joined
Feb 20, 2019
Messages
615
Likes
448
My Earthworks M30 is specified flat from 3 Hz to 30 kHz. The higher model, the M50, is spec'd from 3 Hz to 50 kHz. There are other specialty mics from other companies with significantly higher frequency range (and prices to match).

https://earthworksaudio.com/microphones/m-series/

Yes, I was thinking about the M30, I didn't remember seeing M50 before.

I know, certain DPA models have been very popular for hires recording and also for example Sennheiser 8000-series is good.

This is among the widest band music recording microphones I know, I just don't have it (at least yet):
http://www.sanken-mic.com/en/product/product.cfm/3.1000400
 

andreasmaaan

Master Contributor
Forum Donor
Joined
Jun 19, 2018
Messages
6,652
Likes
9,399
Even 64-bit floating point is not always enough.

Not enough to ensure noise remains below thresholds of audibility? Or not enough to ensure it remains below levels of high-performing analogue gear? Or not enough for what?

By just doing something in digital doesn't mean it is automatically better, because digital has it's own problem domains. Poorly done digital can be easily worse than well done analog.

No argument with that. But poorly-done anything can sound worse than well-done anything else. The discussion here is about well-implemented systems, not poor examples.

None of the DAW's / desks I've seen give out any estimate of remaining calculation precision after chaining certain number of plugins/settings combination.

Ok, but this is a different claim than the claim that any loss of precision results in worse noise and distortion than analogue alternatives. Do you have any evidence for the latter (keeping in mind we are discussing good implementations here, not "old-school" computer music trackers).

And loss of precision (distortion) sounds worse than increase in analog noise floor (noise).

I'd be interested to know if you have any measurements of any high-quality 32-bit floating digital system (hardware or DAW) that demonstrate this?
 

Miska

Addicted to Fun and Learning
Audio Company
Joined
Feb 20, 2019
Messages
615
Likes
448
Not enough to ensure noise remains below thresholds of audibility? Or not enough to ensure it remains below levels of high-performing analogue gear? Or not enough for what?

Yes, not noise, but distortion. Not enough to ensure it remains below levels of high-performing analogue gear. I don't like to make statements about "thresholds of audibility" because that is very complex topic in itself (depends on person, time of the day, mood, listening environment, associated listening gear, etc) and much harder to objectively quantify than performance of analog gear.

I don't know how much you work in audio DSP, but I work on it every day.

No argument with that. But poorly-done anything can sound worse than well-done anything else. The discussion here is about well-implemented systems, not poor examples.

That's where real world meets the theory.

Do you have any evidence for the latter (keeping in mind we are discussing good implementations here, not "old-school" computer music trackers).

Yes, plenty.

But for production side, probably the people who deal with music production daily as work have much more.
 

Miska

Addicted to Fun and Learning
Audio Company
Joined
Feb 20, 2019
Messages
615
Likes
448
Would you care to provide some, please?

Take a DAW, put in bunch of plugins in the chain and turn the parameters to values you know will be difficult for the associated math regarding this aspect, mix let's say 32 channels to one, and run test signals through.

Or alternatively, just run something through MQA pipeline. :D
 

March Audio

Master Contributor
Audio Company
Joined
Mar 1, 2016
Messages
6,378
Likes
9,317
Location
Albany Western Australia
Take a DAW, put in bunch of plugins in the chain and turn the parameters to values you know will be difficult for the associated math regarding this aspect, mix let's say 32 channels to one, and run test signals through.

Or alternatively, just run something through MQA pipeline. :D
Eh?

So put a signal through a bunch of plugins ( I assume you are referring to effects plugins) that distort the sound and you end up with distorted signal.

Not sure what point you are making.

Can you show us the distortion and noise levels of simply mixing 32 channels to 1 as you mentioned.
 
Last edited:

andreasmaaan

Master Contributor
Forum Donor
Joined
Jun 19, 2018
Messages
6,652
Likes
9,399
Take a DAW, put in bunch of plugins in the chain and turn the parameters to values you know will be difficult for the associated math regarding this aspect, mix let's say 32 channels to one, and run test signals through.

You're asking me to provide the evidence to prove your point.

Or alternatively, just run something through MQA pipeline. :D

We're not discussing lossily compressed playback formats here.
 

March Audio

Master Contributor
Audio Company
Joined
Mar 1, 2016
Messages
6,378
Likes
9,317
Location
Albany Western Australia
You're asking me to provide the evidence to prove your point.



We're not discussing lossily compressed playback formats here.
Thats alright, I will (disprove the point) ;)

OK I generated 8 tones at 24 bit resolution -40dBFS at frequencies where their harmonics shouldnt overlap. I loaded them into Adobe Audition and mixed them together into one track (32 bit float) then saved that track as 24 bit with triangular dither.

Result when analysed in Musicscope with a 65k FFT was a noise floor circa -176 dB and no visible harmonics whatsoever.

This is well ahead of the performance of any analogue gear.

1557908949521.png
 
Last edited:

bennetng

Major Contributor
Joined
Nov 15, 2017
Messages
1,634
Likes
1,692
I don't know, but for instance in C# (should I say .net?), limiting the precision in 32-bit float is actually pretty tedious and will result in longer codes. Most math methods and constants are in 64-bit float internally. While a method may return 32-bit float values, the intermediate steps are carried out in higher precision anyway.

Historically, Sonar 5 (2005) has a 64-bit float engine already and Reaper ($60) support 64-bit float as well, it also supports some exotic precision like 39-bit fixed and so on.

I also heard about some older versions of hardware based Pro-Tools running in 48-bit fixed.

So...? No sane people will consider switching to another DAW merely due to mixing precision, the main factors are always other things like user-friendliness, functionality, compatibility and so on.
 

Miska

Addicted to Fun and Learning
Audio Company
Joined
Feb 20, 2019
Messages
615
Likes
448
Thats alright, I will (disprove the point) ;)

OK I generated 8 tones at 24 bit resolution -40dBFS at frequencies where their harmonics shouldnt overlap. I loaded them into Adobe Audition and mixed them together into one track (32 bit float) then saved that track as 24 bit with triangular dither.

Mixing is extremely simplified case of just adding samples together and possibly applying normalization factor. You are trying to way oversimplify the case. This is not representative of typical music production gear or process, from recording of microphone level signals to a final master. Instead add ProTools, typical mixing console, eq, reverb and compression plugins and all other stuff typically part of the process. And run the same process through both analog and digital paths and compare. I can bet Abbey Road or Air Studios don't record and mix in the way you describe in Adobe Audition.

I think it would be already much more representative to compare two of those very popular consoles, 88D (digital) and 88RS (analog/hybrid). Or alternatively SSL AWS or Duality vs some ProTools / Pyramix setup. But especially something like the 5088.

But whatever, nobody of the pro music production world cares what we discuss here, one way or the other. I think they may have a point, and ears, in their domain.
 

Miska

Addicted to Fun and Learning
Audio Company
Joined
Feb 20, 2019
Messages
615
Likes
448
I don't know, but for instance in C# (should I say .net?), limiting the precision in 32-bit float is actually pretty tedious and will result in longer codes. Most math methods and constants are in 64-bit float internally. While a method may return 32-bit float values, the intermediate steps are carried out in higher precision anyway.

So...? No sane people will consider switching to another DAW merely due to mixing precision, the main factors are always other things like user-friendliness, functionality, compatibility and so on.

Many mixing consoles and other digital gear run old school DSP processors, like fixed point TI, Motorola/Freescale 56k-series, or like in case 88D using Analog Devices SHARC.
 

Miska

Addicted to Fun and Learning
Audio Company
Joined
Feb 20, 2019
Messages
615
Likes
448
You're asking me to provide the evidence to prove your point.

You are also asking me to prove your point. Why would I go through the trouble? Instead you can prove your claim that 32-bit floating point will never produce errors larger than what analog gear can represent when used for any audio DSP. And I will have fun reading it.
 

March Audio

Master Contributor
Audio Company
Joined
Mar 1, 2016
Messages
6,378
Likes
9,317
Location
Albany Western Australia
Mixing is extremely simplified case of just adding samples together and possibly applying normalization factor. You are trying to way oversimplify the case. This is not representative of typical music production gear or process, from recording of microphone level signals to a final master. Instead add ProTools, typical mixing console, eq, reverb and compression plugins and all other stuff typically part of the process. And run the same process through both analog and digital paths and compare. I can bet Abbey Road or Air Studios don't record and mix in the way you describe in Adobe Audition.

I think it would be already much more representative to compare two of those very popular consoles, 88D (digital) and 88RS (analog/hybrid). Or alternatively SSL AWS or Duality vs some ProTools / Pyramix setup. But especially something like the 5088.

But whatever, nobody of the pro music production world cares what we discuss here, one way or the other. I think they may have a point, and ears, in their domain.

If you are referring to plugins which by definition distort the signal then you have no way of knowing how effective they are. You are making assumptions and claims you cant back up.

Im not sure on what basis you think that, for example, an analogue compressor may distort more or less than a digital one of the same transfer function. However I see no reason whatsoever that it would be worse than analogue. 32 bit is clearly higher resolution and lower noise than anything analogue.
 
Last edited:

March Audio

Master Contributor
Audio Company
Joined
Mar 1, 2016
Messages
6,378
Likes
9,317
Location
Albany Western Australia
You are also asking me to prove your point. Why would I go through the trouble? Instead you can prove your claim that 32-bit floating point will never produce errors larger than what analog gear can represent when used for any audio DSP. And I will have fun reading it.

Erm because *YOU* made the assertion.
 

March Audio

Master Contributor
Audio Company
Joined
Mar 1, 2016
Messages
6,378
Likes
9,317
Location
Albany Western Australia
You are also asking me to prove your point. Why would I go through the trouble? Instead you can prove your claim that 32-bit floating point will never produce errors larger than what analog gear can represent when used for any audio DSP. And I will have fun reading it.

It was your claim that analogue was better.
 

Miska

Addicted to Fun and Learning
Audio Company
Joined
Feb 20, 2019
Messages
615
Likes
448
It was your claim that analogue was better.

And there was the presumption on the forum before that that music production pro's and companies and wrong and digital is better.

Not my claim, I just said that I understand their point and why that technically can very well be valid.
 

Miska

Addicted to Fun and Learning
Audio Company
Joined
Feb 20, 2019
Messages
615
Likes
448
Erm because *YOU* made the assertion.

Because I know many DSP algorithms where value of 32-bit floating point Epsilon for example will kick your nuts. It is just pure math. Not too hard to hit such cases for example with IIR. And with FIR too. Combine number of such limit cases in the processing pipeline and you errors really begin to grow. And to know the final error amount you need to know all the plugins and all their parameters set up by the user.
 
Status
Not open for further replies.
Top Bottom