Questions About DSD and Audio Sample Rates

[DeDe]

Hi everyone. I’m a beginner student majoring in music production, and my teacher for basic audio tech told us that DSD and higher sample rates in PCM (96k, 192k) can bring advantages in sound quality, but didn’t really go into detail. I did some research on my own and mostly found the following points:

1. Most modern ADCs use a Delta-Sigma architecture (which I don’t fully understand). In the first sampling stage, they use a high‑speed, low‑bit‑depth bitstream similar to DSD, then use a feedback loop to push quantization noise up into very high frequencies, and finally apply digital filtering and decimation to get low‑speed, high‑precision data (for example 24‑bit 48 kHz).

2. Digital filters can be either linear‑phase or minimum‑phase, which respectively introduce ringing and phase distortion (how does this actually present itself in terms of what we hear?).

3. DSD seems to use analog filters to handle the very high‑frequency filtering (what are the benefits of doing it this way?).

4. Higher sample rates can move the Nyquist frequency further above the human hearing range, leaving more room for the steep high‑frequency filters and reducing their distortion.

So I’d like to ask:
– Is DSD audio actually better than typical PCM audio?
– Is there any audible difference between the two formats?
– How big are the differences between different digital filters? And what about in the analog filter domain?
– Do the different sample rates like 44.1, 48, 96 kHz have any real‑world significance?

Also, in EQ plugins you can choose different filter modes. When I switch from zero‑latency mode to linear‑phase mode, apart from the added latency I don’t really hear much difference—at least not on piano tracks.

I’ve seen some audio measurements on forums where they test a device’s ultra‑high‑frequency filters, but what exactly are those graphs showing? Does a lower dB value in the filtered high‑frequency range mean the filter is “better”?

 

[RandomEar]

So I’d like to ask:
– Is DSD audio actually better than typical PCM audio?

No. DSD is a different method to encode information, that's it.

– Is there any audible difference between the two formats?

In general, no. However, DSD tracks might be mastered differently because they appeal to a different audience. And different masterings will sound different. The same is true if you get your PCM tracks from different sources (CD, streaming service A, streaming service B, ...), though.

There is one possible edge case: Trained listeners have shown that for 16 bit audio like on a CD, they managed to hear the difference to 24 bit in controlled testing when listening at elevated levels. The reason - as far as I understand - is that the dithered noise level of 16 bit CD audio is around -96 dB, which can become barely audible if you listen at peak SPLs a good bit above 96 dB. The DSD noise floor is lower than that of 16 bit PCM, so I would imagine that under these extreme conditions, a trained listener could differentiate the two formats during silent passages of a song. With 24 bit PCM, there would be no audible difference to DSD.

– How big are the differences between different digital filters? And what about in the analog filter domain?

I am glad you asked

In short: Between "fast" reconstruction filters, there should be no audible difference. "Slow" filters and NOS mode on average have some treble roll off which can be audible, especially if you're younger. Since that roll off is an inaccuracy, I would not recommend using either slow filters or NOS.

– Do the different sample rates like 44.1, 48, 96 kHz have any real‑world significance?

No. Technically yes: Marketing. But really: No.

As far as we know today, there is no use in replicating ultrasonic sounds above 20 kHz. Therefore, 44.1 kHz sampling is sufficient for humans - at least for music reproduction. For music production, it can be beneficial to have some headroom in sampling rate and bit depth due to unavoidable small losses from filters and format conversions.

Also, in EQ plugins you can choose different filter modes. When I switch from zero‑latency mode to linear‑phase mode, apart from the added latency I don’t really hear much difference—at least not on piano tracks.

Not sure what that is about. Can you link to an example plugin offering that?

Guessing: The plugin may be resampling internally, which means the comments about reconstruction filters from above also apply here. If you need low latency while mixing audio, it's fine to use fast minimum phase filters.

I’ve seen some audio measurements on forums where they test a device’s ultra‑high‑frequency filters, but what exactly are those graphs showing? Does a lower dB value in the filtered high‑frequency range mean the filter is “better”?

You'll have to link such a measurement for us to explain.

 

[DeDe]

No. DSD is a different method to encode information, that's it.


In general, no. However, DSD tracks might be mastered differently because they appeal to a different audience. And different masterings will sound different. The same is true if you get your PCM tracks from different sources (CD, streaming service A, streaming service B, ...), though.

There is one possible edge case: Trained listeners have shown that for 16 bit audio like on a CD, they managed to hear the difference to 24 bit in controlled testing when listening at elevated levels. The reason - as far as I understand - is that the dithered noise level of 16 bit CD audio is around -96 dB, which can become barely audible if you listen at peak SPLs a good bit above 96 dB. The DSD noise floor is lower than that of 16 bit PCM, so I would imagine that under these extreme conditions, a trained listener could differentiate the two formats during silent passages of a song. With 24 bit PCM, there would be no audible difference to DSD.


I am glad you asked

In short: Between "fast" reconstruction filters, there should be no audible difference. "Slow" filters and NOS mode on average have some treble roll off which can be audible, especially if you're younger. Since that roll off is an inaccuracy, I would not recommend using either slow filters or NOS.


No. Technically yes: Marketing. But really: No.

As far as we know today, there is no use in replicating ultrasonic sounds above 20 kHz. Therefore, 44.1 kHz sampling is sufficient for humans - at least for music reproduction. For music production, it can be beneficial to have some headroom in sampling rate and bit depth due to unavoidable small losses from filters and format conversions.


Not sure what that is about. Can you link to an example plugin offering that?

Guessing: The plugin may be resampling internally, which means the comments about reconstruction filters from above also apply here. If you need low latency while mixing audio, it's fine to use fast minimum phase filters.


You'll have to link such a measurement for us to explain.

Thanks for your reply! That post of yours pretty much completely cleared up my questions about digital filters.

You also reminded me of something. I went and found the FabFilter manual, where they explain the three processing modes of their EQ plugin (besides the two you already mentioned, there’s also a “Natural Phase” mode). They also provide specific frequency/phase graphs and demo videos. According to the graphs, in the frequency domain they only differ in the very top end above 18 kHz. But other than a linear-phase filter, they all affect phase, while linear-phase mode introduces pre-ringing, which matches what I’d understood before.

They also mention that linear-phase filters are only recommended in a few special cases, like a steep low-pass or high-pass filter, or when you’re playing processed and unprocessed versions of a similar signal at the same time and that would cause cancellation. Otherwise, non-linear-phase filters like Zero Latency mode and “Natural Phase” mode are what you’d use in most situations.

In FabFilter’s demo video, they specifically use a steep high-pass filter to compare Zero Latency mode and Linear Phase mode. With the high-pass, Zero Latency mode shows a 180° phase flip, while the linear-phase filter’s pre-ringing really does show up in the low end as a sort of “reverb that happens early,” which is pretty wild.

But that’s a filter used for music production, and it’s an extreme case. Ringing only happens around the frequencies where the filter is acting, right? Like you said in your post, the DAC’s linear-phase filter causes pre-ringing at 22.05 kHz, which people basically can’t hear.

About measurement—like the filter measurement data in this post：https://www.audiosciencereview.com/...ds/topping-e2x2-audio-interface-review.49127/

 

[RandomEar]

[...]

But that’s a filter used for music production, and it’s an extreme case. Ringing only happens around the frequencies where the filter is acting, right? Like you said in your post, the DAC’s linear-phase filter causes pre-ringing at 22.05 kHz, which people basically can’t hear.

For linear phase filters at 44.1 kHz, yes. For 48 kHz it would be ringing at 24 kHz (= Nyquist frequency) and so on. Minimum phase filter ringing seems to be at a slightly lower frequency, but also shouldn't be a concern.

About measurement—like the filter measurement data in this post：https://www.audiosciencereview.com/...ds/topping-e2x2-audio-interface-review.49127/

OK, so this graph?

It shows the filter attenuation over the frequency. So 0 dB is "no attenuation at all", -20 dB is 90% attenuation, -40 dB is 99% attenuation and so on. Essentially, you would want perfect attenuation above half the sampling rate (so 22.05 kHz for this example) and no attenuation at all below that frequency. Sadly, infinitely sharp filters don't exist and keeping all other qualities identical, sharper filters need more "taps" and introduce more delay. In practice, filters are therefore usually designed to not attenuate below 20 kHz and to reach the full attenuation of about -100 dB by roughly 24 kHz (for 44.1 kHz material). This specific filter in the Topping E2x2 is lacking some attenuation.

 

[dfuller]

– Is DSD audio actually better than typical PCM audio?

No, just different. In fact in a music production sense it's way worse since you can't process DSD - including something as simple as fades.

– Is there any audible difference between the two formats?

Not really, no.

– How big are the differences between different digital filters? And what about in the analog filter domain?

Well, it depends, doesn't it? Most linear phase fast roll-off anti-aliasing filters sound the same - It is plausible that better filter designs with more taps could be better technically, but whether or not it's audible is a different question. Slow roll-off and NOS (non-oversampled) filters will probably make an audible difference and not for the better.

– Do the different sample rates like 44.1, 48, 96 kHz have any real‑world significance?

They can, with certain plugins. If the plugins aren't doing a proper job of oversampling, a higher sample rate will certainly sound different from a lower one. Soundtoys is the most obvious here.

Also, in EQ plugins you can choose different filter modes. When I switch from zero‑latency mode to linear‑phase mode, apart from the added latency I don’t really hear much difference—at least not on piano tracks.

Switching from IIR minimum phase ("zero latency") to FIR linear phase ("linear phase") can be audible with the right filters - these are usually very steep and at low frequencies.

 

[danadam]

In the first sampling stage, they use a high‑speed, low‑bit‑depth bitstream similar to DSD, then use a feedback loop to push quantization noise up into very high frequencies

AFAIK not "then" but "during". You can only "move" noise when it is created and after that it's baked-in in the signal.

3. DSD seems to use analog filters to handle the very high‑frequency filtering (what are the benefits of doing it this way?).

Your marketing team can spin it as more "natural"

Do the different sample rates like 44.1, 48, 96 kHz have any real‑world significance?

In playback not really. In production they are used by effects that do non-linear operations to reduce aliasing resulting from those operations. It's nicely described in this video by Dan Worrall: