Help with understanding sampling rate for DSP.

[Oski1928]

Hi, quick question. I’m looking at the DMP-A8 and its DSP capabilities. I see people talking about the fact that the DSP resamples hi res audio to 48khz. On this site I thought it was pretty well established that “hi res” audio (specifically its sampling rate for the purposes of this discussion) has no audible benefit. If this is the case then why would the same group of people view fact that DSP resamples to 48khz as a negative?

I have no opinion here and actually ask as a question. Am I missing something in regards to sampling rate being different in this context? Thanks!

 

[Keith_W]

The higher the sampling rate, the finer your corrections can be, but at a price.

Suppose you want to time align speaker A to speaker B. It could be your midrange to your tweeter, or surround speakers to listening position, or mains to subs. At 48kHz sampling rate, you can do it in (1/48000 * 1000) = 0.0208ms increments. At 96kHz sampling rate, that's 0.0104ms. I highly doubt anybody can hear that difference, but it does produce nice looking graphs to show off on ASR

The downside of higher sampling rate is that you lose resolution if you are limited by processing power. Some less powerful DSP units might have 1024 taps per channel. So at 48kHz, the frequency "block" you can correct is 48000/1024 = 46.9Hz. If you sample at 96kHz, it's 93.8Hz. Sampling at 48kHz with only 1024 taps only gives you 2 bins to correct below 100Hz, that is clearly not enough. In that case you are better off sampling at 24kHz if your DSP software/hardware allows it, since it doubles the number of bins available to 4. Still inadequate IMO but better than only 2.

 

[Oski1928]

The higher the sampling rate, the finer your corrections can be, but at a price.

Suppose you want to time align speaker A to speaker B. It could be your midrange to your tweeter, or surround speakers to listening position, or mains to subs. At 48kHz sampling rate, you can do it in (1/48000 * 1000) = 0.0208ms increments. At 96kHz sampling rate, that's 0.0104ms. I highly doubt anybody can hear that difference, but it does produce nice looking graphs to show off on ASR

View attachment 340172

The downside of higher sampling rate is that you lose resolution if you are limited by processing power. Some less powerful DSP units might have 1024 taps per channel. So at 48kHz, the frequency "block" you can correct is 48000/1024 = 46.9Hz. If you sample at 96kHz, it's 93.8Hz. Sampling at 48kHz with only 1024 taps only gives you 2 bins to correct below 100Hz, that is clearly not enough. In that case you are better off sampling at 24kHz if your DSP software/hardware allows it, since it doubles the number of bins available to 4. Still inadequate IMO but better than only 2.

Lost me a little with bins, but overall thank you for the information. I definitely understand your initial and main point.

Edit: I put a laughing emoji after “bins”. I guess that doesn’t translate to the forum, but I wanted to provide context to the tone of my statement.

 

[Keith_W]

Umm, the last time I used this analogy on ASR I was criticized for it ... but it does make it easier to understand. Think of a "bins" as bands in a parametric equalizer. Suppose you have a 10 band PEQ and you set it up so that it only corrects the lowest 1000Hz in your system. Each band will be 100Hz wide. If you make it correct 10kHz, then each band is 1000Hz wide.

In the same way, 1024 "bands" in a 48kHz system = each band is 46.9Hz wide ... 1024 bands in a 96kHz system = each band 93.8Hz ... and so on.

More powerful DSP units have more "taps", e.g. I use Acourate which gives you 131,000 taps per channel. So if I sample at 48kHz, my "bands" are only 0.37Hz wide. If I want to correct below 100Hz, I have (100/0.37) = 270 bins to play with. This is clearly overkill and most of us (or maybe all of us!) don't need this, but it is nice to have.

 

[antcollinet]

At 48kHz sampling rate, you can do it in (1/48000 * 1000) = 0.0208ms increments

I think that is not the case. With Digital sampling there is no time resolution problem - as shown in the monty video (shortly after 20:30). You can time shift the band limited signal by very much smaller amounts*, independent of sample rate.

*See post #10 from @pkane below.


EDITED : to correct time resolution from infinte.

 

[Oski1928]

Umm, the last time I used this analogy on ASR I was criticized for it ... but it does make it easier to understand. Think of a "bins" as bands in a parametric equalizer. Suppose you have a 10 band PEQ and you set it up so that it only corrects the lowest 1000Hz in your system. Each band will be 100Hz wide. If you make it correct 10kHz, then each band is 1000Hz wide.

In the same way, 1024 "bands" in a 48kHz system = each band is 46.9Hz wide ... 1024 bands in a 96kHz system = each band 93.8Hz ... and so on.

More powerful DSP units have more "taps", e.g. I use Acourate which gives you 131,000 taps per channel. So if I sample at 48kHz, my "bands" are only 0.37Hz wide. If I want to correct below 100Hz, I have (100/0.37) = 270 bins to play with. This is clearly overkill and most of us (or maybe all of us!) don't need this, but it is nice to have.

Thanks for that, helped give me a better understanding for sure!

 

[Keith_W]

I think that is not the case. With Digital sampling there is no time resolution problem - as shown in the monty video (shortly after 20:30). You can time shift the band limited signal by infinitely small amounts regardless of sample rate.

Yeah, aware of that. I didn't want to make the explanation too complicated.

BTW I haven't seen that Monty video ... can I have a link, please?

 

[antcollinet]

BTW I haven't seen that Monty video ... can I have a link, please?

Oh I love the Monty video. Any excuse to share

 

[TurtlePaul]

In that case you are better off sampling at 24kHz if your DSP software/hardware allows it, …

Most people can hear more than 12 khz, so sampling of at least 44 khz is recommended.

 

[pkane]

I think that is not the case. With Digital sampling there is no time resolution problem - as shown in the monty video (shortly after 20:30). You can time shift the band limited signal by infinitely small amounts regardless of sample rate.

There is a limit to time resolution of sampled audio, but it isn't related to the sampling rate, but rather to the number of bits per sample:

Time resolution of digital audio

A frequent claim by detractors of digital audio is that the time resolution is equal to the sampling interval, 22.7 μs for the CD format. This is incorrect. Although there is a limit, it is much sm…

troll-audio.com

 

[pkane]

Hi, quick question. I’m looking at the DMP-A8 and its DSP capabilities. I see people talking about the fact that the DSP resamples hi res audio to 48khz. On this site I thought it was pretty well established that “hi res” audio (specifically its sampling rate for the purposes of this discussion) has no audible benefit. If this is the case then why would the same group of people view fact that DSP resamples to 48khz as a negative?

I have no opinion here and actually ask as a question. Am I missing something in regards to sampling rate being different in this context? Thanks!

Resampling is just an extra step in the process that can be done well, or not so well. Resampling to the same rate makes the job of the DSP implementer easier, but otherwise, doesn't necessarily improve anything. It can add distortions and other artefacts if not done correctly.

I suspect that "purists" will insist on no resampling, but that isn't something that can be claimed to be better in general. It'll depend on the actual DSP+resampler implementation. Some DSP may actually need upsampling internally to eliminate aliases and to simplify filtering.

 

[antcollinet]

There is a limit to time resolution of sampled audio, but it isn't related to the sampling rate, but rather to the number of bits per sample:

Time resolution of digital audio

A frequent claim by detractors of digital audio is that the time resolution is equal to the sampling interval, 22.7 μs for the CD format. This is incorrect. Although there is a limit, it is much sm…

troll-audio.com

Thanks for the clarification. I think I've been aware of this in the past, and my "infinite" statement is clearly incorrect.

Though we should also point out : nearly 1000 times better at red-book. Nearly 256000 times better at 24bit.

I'm not going to lose sleep any time soon - though, for accuracy I'll remove "infinitely" from my post.

 

[voodooless]

Umm, the last time I used this analogy on ASR I was criticized for it ... but it does make it easier to understand. Think of a "bins" as bands in a parametric equalizer. Suppose you have a 10 band PEQ and you set it up so that it only corrects the lowest 1000Hz in your system. Each band will be 100Hz wide. If you make it correct 10kHz, then each band is 1000Hz wide.

In the same way, 1024 "bands" in a 48kHz system = each band is 46.9Hz wide ... 1024 bands in a 96kHz system = each band 93.8Hz ... and so on.

More powerful DSP units have more "taps", e.g. I use Acourate which gives you 131,000 taps per channel. So if I sample at 48kHz, my "bands" are only 0.37Hz wide. If I want to correct below 100Hz, I have (100/0.37) = 270 bins to play with. This is clearly overkill and most of us (or maybe all of us!) don't need this, but it is nice to have.

To be clear, this only applies to (certain types of) convolution filters (FIR). The majority of DSP solutions commercially available do not support FIR. The alternative, IIR does not have such limitations and its used in the vast majority of cases. A hybrid of both is also possible, like Dirac or Audyssey do.

 

[antcollinet]

Most people can hear more than 12 khz, so sampling of at least 44 khz is recommended.

I think that was a typo. It was intended to be double 48KHz.

 

[Deleted member 21114]

Oh I love the Monty video. Any excuse to share

The First Monty of 2024! Happy New Year y'all!

 

[MCH]

The First Monty of 2024! Happy New Year y'all!

I see that Monty, and raise a Youngmoo Kim (more obscure but i highly recommend it - it is a series)

 

[Multicore]

Oh I love the Monty video. Any excuse to share

Yeah. It's a good one. I'm very suspicious of the overall genre called YouTube Splainer and often have allergic reactions when exposed to them but this one is really good.

 

[Multicore]

I see that Monty, and raise a Youngmoo Kim (more obscure but i highly recommend it - it is a series)

i got an allergic reaction from the gratuitous background music and visual effects. and the pronouncement "A signal is anything that provides information or data." i'll take that sentence outside and beat it up. 1. information is not the same as data. 2. a research intern can provide data but i wouldn't call her a signal.

YouTube Splainers. tough genre.

 

[MCH]

i'll take your sentences outside and beat them up.

1. information is not the same as data.

and the sentence does not say that, more the opposite, it says "information OR data"

2. a research intern can provide data but i wouldn't call her a signal.

neither says that, it says that a signal can provide data, not that any data comes always in the form of a signal

YouTube Splainers. tough genre.

It is not a YouTube Splainer. As written in the description is a course from the Drexel University, which i suspect but can't assure, professor Kim recording during covid times. I really recommend you to watch the series, it is very easy to follow and the best explanation for noobs that i know of.