Upsampling with cheap chips discardes some of the good samples and produces fake samples, based on some "coeficients" stored internally in the ROM.
The DSP aproach, because of the generous ammount of RAM, can maintain all the initial samples and generate the intermediate ones based on calculations done in real time.
The DSP aproach, because of the generous ammount of RAM, can maintain all the initial samples and generate the intermediate ones based on calculations done in real time.
Worked on few Denon CD and DVD players with DXP6001AF....noisy mf... Denon implementation is even noisier... quite appalling..... unbelievable would be the better word.
Boky
Last edited:
What a load of nonsense! The drawbacks of 'cheap chips' vs DSP are down to the implementation differences. 'Fake' samples will always be produced if the FIR filter isn't a half-band one - that's the only kind which preserves the original samples when upsampling. And - as it turns out - most of the the 'cheap chips' do indeed use half-band filters. One of the advantages of going to DSP is to get away from such limitations - half-band filters by definition allow aliasing and hence violate one of the prime directives of digital audio.
Hi,
I do not where you get your information from, but it is plain and simple false.
First, Asynchronous Upsampling, no matter technology is used, MUST re-calculate each and every sample, there is no other way, as original samples and output samples do not line up.
No matter how the intermediate samples are calculated (straightline interpolation, various filter functions etc.) the sample rate converted samples MUST be different from those that where originally present and will contain errors who's magnitude and nature depends how closely sample rate conversion process and signal match each other, or not.
But no matter what, if the sample rate conversion is asynchronous, ALL output samples are "Fake" and no "Real" samples are retained.
Any possible retention of "real samples" could only happen if the upsampling operates with integer ratio's, that is, if we use the special case of upsampling usually called oversampling. This would however only possible if we do not apply a filter during the oversampling process, if a digital filter is employed then again, all samples output by the filter are "Fake".
The above holds equally true if the algorythms are impelemented in a hardware chip (be it an SAA7220 or DF1704 for oversampling or an AD1896 for ASRC) or if they are running in software on a Quad Core i7 with 16GB or RAM.
So, in fact the only way to preserve the "Real Samples" would be to avoid any digital manipulation, that is to operate "Non-Oversampling". If this "preserving real samples" is a goal that meaningful is open to debate.
Ciao T
I do not where you get your information from, but it is plain and simple false.
First, Asynchronous Upsampling, no matter technology is used, MUST re-calculate each and every sample, there is no other way, as original samples and output samples do not line up.
No matter how the intermediate samples are calculated (straightline interpolation, various filter functions etc.) the sample rate converted samples MUST be different from those that where originally present and will contain errors who's magnitude and nature depends how closely sample rate conversion process and signal match each other, or not.
But no matter what, if the sample rate conversion is asynchronous, ALL output samples are "Fake" and no "Real" samples are retained.
Any possible retention of "real samples" could only happen if the upsampling operates with integer ratio's, that is, if we use the special case of upsampling usually called oversampling. This would however only possible if we do not apply a filter during the oversampling process, if a digital filter is employed then again, all samples output by the filter are "Fake".
The above holds equally true if the algorythms are impelemented in a hardware chip (be it an SAA7220 or DF1704 for oversampling or an AD1896 for ASRC) or if they are running in software on a Quad Core i7 with 16GB or RAM.
So, in fact the only way to preserve the "Real Samples" would be to avoid any digital manipulation, that is to operate "Non-Oversampling". If this "preserving real samples" is a goal that meaningful is open to debate.
Ciao T
Hi,
Many modern DAC IC's include Digital filters and even ASRC on board. It is rather hard to separate out the "sound of the DAC" from "the sound of the Filter" as a result...
Ciao T
Many modern DAC IC's include Digital filters and even ASRC on board. It is rather hard to separate out the "sound of the DAC" from "the sound of the Filter" as a result...
Ciao T
Yeah, I know the story, NOS all the way is the true light.
Thank you, my bad, AD1896...
Folks,
For those interested in what different filters do, the March Stereophile will feature a review of a Digital Product that offers a very wide selection of digital filters, including "no filters at all" and multiple DAC's.
Many (but not all) of the filters are documented with detailed measurements, though sadly the reviewer did not spend a lot time on the sonic differences, so it is not as interesting as it could be.
During the development of this product I found myself (as well as many others) subjectively preferring minimum phase filters with no pre-ringing preferable to traditional "brickwall - symmetrical impulse response" types, even under blind conditions.
What surprised me however was that I did however NOT prefer the filter of these that was most like Non-Os (that would be one with no pre-ringing and only four cycles of visible post-ringing) but one that is essentially more like the classic LC Brickwall filter on early non-oversampling ADC/DAC Systems (Sony PCM F-1 for example) and first generation Japanese CD-Players, which has 19 cycles of visible post-ringing.
However, in this particular area the preference for the "short"and "long" minimum phase filter was split with no particular pattern of preference notable
Generally Non-Os with some analogue filtering came out on top, followed by "unfiltered" Non-Os and the analogue-like minimum phase filters. Different versions of traditional symmetrical impulse response digital filters where ranked at the bottom and found to be very little different from each other.
Ciao T
For those interested in what different filters do, the March Stereophile will feature a review of a Digital Product that offers a very wide selection of digital filters, including "no filters at all" and multiple DAC's.
Many (but not all) of the filters are documented with detailed measurements, though sadly the reviewer did not spend a lot time on the sonic differences, so it is not as interesting as it could be.
During the development of this product I found myself (as well as many others) subjectively preferring minimum phase filters with no pre-ringing preferable to traditional "brickwall - symmetrical impulse response" types, even under blind conditions.
What surprised me however was that I did however NOT prefer the filter of these that was most like Non-Os (that would be one with no pre-ringing and only four cycles of visible post-ringing) but one that is essentially more like the classic LC Brickwall filter on early non-oversampling ADC/DAC Systems (Sony PCM F-1 for example) and first generation Japanese CD-Players, which has 19 cycles of visible post-ringing.
However, in this particular area the preference for the "short"and "long" minimum phase filter was split with no particular pattern of preference notable
Generally Non-Os with some analogue filtering came out on top, followed by "unfiltered" Non-Os and the analogue-like minimum phase filters. Different versions of traditional symmetrical impulse response digital filters where ranked at the bottom and found to be very little different from each other.
Ciao T
Seeing as half-band FIR doesn't suit me as regards technical performance, no thanks, no need to go there. What are you basing your claim on that it doesn't incidentally? I may have misinterpreted the context, I wasn't speaking of async SRC, rather (fixed) integer ratio upsampling. I'm glad to be corrected if you have a reference.
Hi,
It is the way the filters are build, they invariably include more than two samples, the output must re-calculate all samples to work.
I suspect one could write an algorithm that explicitly preserves the original samples, but I do not know any.
Ciao T
Seeing as half-band FIR doesn't suit me as regards technical performance, no thanks, no need to go there. What are you basing your claim on that it doesn't incidentally? I may have misinterpreted the context, I wasn't speaking of async SRC, rather (fixed) integer ratio upsampling. I'm glad to be corrected if you have a reference.
It is the way the filters are build, they invariably include more than two samples, the output must re-calculate all samples to work.
I suspect one could write an algorithm that explicitly preserves the original samples, but I do not know any.
Ciao T
By your original response I thought it was possible that you'd tried it. After all, you did tell me to. So no references, no empirical results?
Hi,
I can only talk about the filters I encountered, even the halve band ones, with integer oversampling do not preserve samples, because of their fundamental design and based on observations of their output (sorry, nothing published - I research things for my own knowledge, not to publish - there are enough academics to do the latter).
And no, I have not tried to generate a filter algorithm that preserves the original samples, nor have I turned every conceivable stone to see if one lurks under one. Hence my point of "possible, but never encountered".
Ciao T
I can only talk about the filters I encountered, even the halve band ones, with integer oversampling do not preserve samples, because of their fundamental design and based on observations of their output (sorry, nothing published - I research things for my own knowledge, not to publish - there are enough academics to do the latter).
And no, I have not tried to generate a filter algorithm that preserves the original samples, nor have I turned every conceivable stone to see if one lurks under one. Hence my point of "possible, but never encountered".
Ciao T
Hi ThorstenL,
Thanks for this useful summary.
I remember this Dirac impulse response. I find something very similar in the Philips CD 723. (See attached picture).
Attachments
Hi,
Yes, this is broadly what it looks like.
No wonder I really liked my heavily modded Philips CD-720 (very similar design, slightly different chipset but same digital filter) so much. I never tested impulse response on that one...
BTW, I still liked true Non-OS better than that...
Ciao T
Yes, this is broadly what it looks like.
No wonder I really liked my heavily modded Philips CD-720 (very similar design, slightly different chipset but same digital filter) so much. I never tested impulse response on that one...
BTW, I still liked true Non-OS better than that...
Ciao T
Well here's a hand-waving kind of argument to demonstrate that you're mistaken.
Let's take a simple interpolate-by-two function achieved with a half-band filter. These filters have the interesting property that all even-numbered coefficients are precisely zero, and the central coefficient we shall assume to be unity.
Therefore the only contribution to every other sample generated is precisely one of the input samples multiplied by the central coefficient, which is unity. All other contributions being multiplied by zero are zero. Thereby the original input sample is preserved. In between the original input samples of course we have the convolution with the various non-zero coefficients.
That is correct... for plain OS. Of course that is just a theoretical example that will do nothing in the real world, but yes, one can keep the original samples if does integers and clock is the same for input/output.
As far as I know only the off-the-shelf ASRC filter cips are not doing that, because the input and output frequency domains are separated (even if equal). So they will have slight differences in frequency/phase that, without serious buffering, need to be dealt with "fake" smples.
As far as I know only the off-the-shelf ASRC filter cips are not doing that, because the input and output frequency domains are separated (even if equal). So they will have slight differences in frequency/phase that, without serious buffering, need to be dealt with "fake" smples.
When you look at such impulse responses please look at them in spectral view. You´ll see that this pre and post ringing only is happening above the cutoff frequency of the lowpass-filter. Most likely no one can hear this. If you use filters that change the content well below 20kHz you may of cause hear differences. A gentle linear lowpass with defined low aliasing may be best.
As for upsampling. Upsampling needs some lowpassing so the resulting data will have 100% different samples to the original if the music has content/noise up to the cutoff frequency of the upsampler. This should be valid for most music.
We lately had a thread at Hydrogenaudio that showed that with the sox resampler it is even possible to keep samples intact while upsampling with some synthetic samples that have no contenet near the cutoff frequency.
Half-Band filters seem to be a way but these indeed add aliasing that reflects back in the music and most likely can be heard.
Sample rate conversion - Hydrogenaudio Forums
Last edited:
Were these linear phase filters by any chance designed to be equi-ripple? If so what was the passband ripple figure?
This is right. But keep in mind, that the term "DAC" resp. digital to analoge converter is ambiguous and can mean both the DAC-IC itself (e. g. PCM63, PCM1702, PCM1704, AD1955) and a complete device (e. g. XiangSheng DAC-02A, Cambridge Audio DAC3/"DAC-3", YULONG DAH1 or Parasound dac 1000) respective a DAC PCB section inside of the cd player.
Unfortunately I don't mentioned exact, what I mean of this both.
Most members does assumed, that I mean the first mentioned. However - I mean the second, thus include digital filter section, S-P/DIF receiver, reclocking approaches and upsampling/oversampling approaches.
Subject of the discussion should actually be the "front end" ICs, which I mentioned by post #1.
It would have been much clearer, if I had call the issue of the headline "... before digital audio interface" (i. e. before I2S resp.Sony-Burr-Brown/SPDIF Interface) instead "....before DAC")
Last edited:
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.