A new thread, I tried the glove and it fits.
I hope people can give their experiences with non os DAC's, listening experiences with known brand non os DAC's ( Audio Note etc. ), which chips to choose etc. and share them with others.
Also people that don't like non os are welcome to give their opinions and technical explanations.
I like both techniques to be clear.
Please let the thread be respectful to adepts of both techniques.
Jean-Paul
I hope people can give their experiences with non os DAC's, listening experiences with known brand non os DAC's ( Audio Note etc. ), which chips to choose etc. and share them with others.
Also people that don't like non os are welcome to give their opinions and technical explanations.
I like both techniques to be clear.
Please let the thread be respectful to adepts of both techniques.
Jean-Paul
My Denon DCD-1000 from 1985 with Burr-Brown PCM56 sounds crap...if you compare to a modern player of rather cheap quality. My DCD-1500 from 1986 with 2x oversampling sounds also crap but with slightly better than DCD-1000. DCD-1000 has a LC filter with 7th order and DCD-1500 has a 9th order LC filter. My point of view is clear. NON-OS are history! Nothing can change my mind. Tubes has a certian character but "old" digital sound is just cold.
I don't know the DCD 1000 but does it have one or two PCM56's ?
I personally never liked the PCM56 and remember replacing them with AD1860 gave a big difference in quality of signal.
Newer non os DAC's don't have the higher order filters ( if they have filtering at all ). Could the way they are filtered not be the reason for the cold sound you describe...
Jean-Paul
I personally never liked the PCM56 and remember replacing them with AD1860 gave a big difference in quality of signal.
Newer non os DAC's don't have the higher order filters ( if they have filtering at all ). Could the way they are filtered not be the reason for the cold sound you describe...
Jean-Paul
The first CD has only one DAC with sample and hold in order to not create a time differnce. The second one has two DAC's.
Filters are central in digital reproduction. You don't have them because it's fun, it's neccesary. And yes, LC-filters are poor, that's why they came up with oversampling in order to get rid of them!. Nothing beats digital filters, can be made as good as you want. You forget one thing: In Europe it's forebidden to have equipment which generates much RF emission. This is a side effect but not unimportant.
Filters are central in digital reproduction. You don't have them because it's fun, it's neccesary. And yes, LC-filters are poor, that's why they came up with oversampling in order to get rid of them!. Nothing beats digital filters, can be made as good as you want. You forget one thing: In Europe it's forebidden to have equipment which generates much RF emission. This is a side effect but not unimportant.
How about this then...
Ideally we would record with non-oversampling Flash or SAR ADCs running at 192kHz or more, and we would replay with non-oversampling R2R DACs running at the same rate.
Please think this thoroughly through before attempting to refute this.
About the filters. They are necessary only if you subscribe to the notion that reconstruction filtering is required. Well, it is required if you want to recover the original waveform. But since we are looking at 44.1k sampling (at least), and since we accept that the ear cuts off sharply at 20k or lower, the necessity for reconstruction seems a lot less strict.
Digital filters are not a cure at all. The one function we really need here, Sinc, is a function they can not implement with finite means. The approximation gets better with more oversampling, but please then show me a 20 bit R2R DAC that can keep up with, say, 1024 times oversampling.
See where this leads at?
Oversampling DACs are wrong for one specific reason (no filter approximates Sinc like we want to).
Non-oversampling DACs are wrong for one reason (HF injection into the system), and may be wrong for a second reason (reconstruction required or not).
Ideally we would record with non-oversampling Flash or SAR ADCs running at 192kHz or more, and we would replay with non-oversampling R2R DACs running at the same rate.
Please think this thoroughly through before attempting to refute this.
About the filters. They are necessary only if you subscribe to the notion that reconstruction filtering is required. Well, it is required if you want to recover the original waveform. But since we are looking at 44.1k sampling (at least), and since we accept that the ear cuts off sharply at 20k or lower, the necessity for reconstruction seems a lot less strict.
Digital filters are not a cure at all. The one function we really need here, Sinc, is a function they can not implement with finite means. The approximation gets better with more oversampling, but please then show me a 20 bit R2R DAC that can keep up with, say, 1024 times oversampling.
See where this leads at?
Oversampling DACs are wrong for one specific reason (no filter approximates Sinc like we want to).
Non-oversampling DACs are wrong for one reason (HF injection into the system), and may be wrong for a second reason (reconstruction required or not).
First of all, the analog output stage is (a major) part of the equation and should be noted. Ears are analog.
TDA1541 + SAA7220 + opamp: dry sound.
TDA1541 + SAA7220 + resistor IV + passive filter + tube output driver: more spatial, lacking bass.
TDA1541 + SAA7220 + tube IV + passive filter + tube output: bass better than before, more linear, more dynamic
TDA1541 + SAA7220 + bjt IV + passive filter + tube output:
best dynamics with a 1541
TDA1541 + SAA7220 + bjt IV + passive filter + emitter follower output:
only 1% less than previous.
TDA 1541 non-os + passive output: slow, muddy, dark.
TDA 1541 non-os + tube output: not so dark anymore, but harsh.
PCM63 non-os + tube output: less harsh than 1541, but not relaxed.
PCM63 + SM5842 + opamp output: details are good, but sounds restless and nervous. Not good for longterm listening.
PCM63 + SM5842 + tube output: more air, better dynamics and detail. Still not so smooth.
PCM63 + PMD100 + tube output: best multibit system. Not as smooth as 1-bit, very up front stage. Tried with PCM56, PCM58, PCM61, PCM 1701, 1702 and 1704 with little differences.
Pioneer LegatoLink + tube output: space, air, smooth and easy to the ear. Details aren't en par with pcm63.
PCM1732 (sigma/delta) + tube output: BB sound (reminds of PCM6x series) with 1 bit smoothness and relaxedness.
TDA1305 (sigma/delta) + tube output: muddy bass, but very spatial.
SAA7325, SAA7350, TDA1547 + tube output: what a mess.
AKM EK4357 + opamp output: Highs a bit harsh, especially in PCM mode. Long warm-up time.
AKM EK4357 + tube output: very balanced, especially nice in the bass with DSD. No harshness anymore.
AKM EK4357 + transistor output: harshness is less, warm-up time remains.
BB DSD1700 + tube output: little lift in the highs, lacking bass. Not too spatial.
I tried a variety of other dacs, but most weren't in my system for very long, so a good judgement can't be given.
Remco
TDA1541 + SAA7220 + opamp: dry sound.
TDA1541 + SAA7220 + resistor IV + passive filter + tube output driver: more spatial, lacking bass.
TDA1541 + SAA7220 + tube IV + passive filter + tube output: bass better than before, more linear, more dynamic
TDA1541 + SAA7220 + bjt IV + passive filter + tube output:
best dynamics with a 1541
TDA1541 + SAA7220 + bjt IV + passive filter + emitter follower output:
only 1% less than previous.
TDA 1541 non-os + passive output: slow, muddy, dark.
TDA 1541 non-os + tube output: not so dark anymore, but harsh.
PCM63 non-os + tube output: less harsh than 1541, but not relaxed.
PCM63 + SM5842 + opamp output: details are good, but sounds restless and nervous. Not good for longterm listening.
PCM63 + SM5842 + tube output: more air, better dynamics and detail. Still not so smooth.
PCM63 + PMD100 + tube output: best multibit system. Not as smooth as 1-bit, very up front stage. Tried with PCM56, PCM58, PCM61, PCM 1701, 1702 and 1704 with little differences.
Pioneer LegatoLink + tube output: space, air, smooth and easy to the ear. Details aren't en par with pcm63.
PCM1732 (sigma/delta) + tube output: BB sound (reminds of PCM6x series) with 1 bit smoothness and relaxedness.
TDA1305 (sigma/delta) + tube output: muddy bass, but very spatial.
SAA7325, SAA7350, TDA1547 + tube output: what a mess.
AKM EK4357 + opamp output: Highs a bit harsh, especially in PCM mode. Long warm-up time.
AKM EK4357 + tube output: very balanced, especially nice in the bass with DSD. No harshness anymore.
AKM EK4357 + transistor output: harshness is less, warm-up time remains.
BB DSD1700 + tube output: little lift in the highs, lacking bass. Not too spatial.
I tried a variety of other dacs, but most weren't in my system for very long, so a good judgement can't be given.
Remco
Unrefuted. I'd even go for 2.8 Mhz.
If only it were as simple as that. However, the output signal with all its HF passes through a series of other circuits, often including opamps with 120 dB of gain and 100 dB of feedback, and a (generally observed) higly non-linear element: the loudspeaker.
In all of these circuits intermodulation occurs and even the intermodulation products may intermodulate again in a subsequent circuit.
This accumulation of IMD is like poison in a food chain: the last one gets the most. In this case, the loudspeaker. And the listener.
Remco
You can not look at the issue on Non-OS DACS with little or no reconstruction filter without looking at everything that is connected to the output of your DAC, i.e. your amp, speakers, and ears!. While your ears can not hears any sounds over 20KHz, if anything in the reproduction chain has enough IM distortion to create artifacts under 20KHz, then you will hear it. That could be your amp, your speakers, and yes, even your ears, though I doubt they would be very effective at creating IM distortion....
Most oversampling DAC implementations use off the shelf oversampling filters. As one other poster has indicated, most off the shelf oversamplers are based on symetric linear phase FIR filters. These exhibit ringing that many consider not to sound good. There is no reason why you could not implement a digital filter that does not ring. You can implement a bessel filter in the digital domain [or you could use a minimum phase FIR filter].
Often when people create digital representations of analog filters, they use a zero order hold approximation [as Werner alludes to]. The zero order hold essentially assumes the signal is the same from the time it was sampled until the time the next sample is taken..... i.e. the typical flat-top digital representation of a signal. A way to improve this approximation, as Werner says, is to increase the sampling frequency. In my old days of designing control systems (one of many hats), we used to say that the zero order hold was only a good approximation if the sample frequency was 10 times the frequencies of interest....
However, higher sampling frequency is not the only way to improve the approximation. You can use higher order holds to create digital approximations of analog filters.
.... sorry if the theory is going over your head [I don't know your background], but this argument of non-os vs. os DACS I find is frought with a lack of understanding of the issues at hand or based on reviews of products that are not ideally implemented.
Most oversampling DAC implementations use off the shelf oversampling filters. As one other poster has indicated, most off the shelf oversamplers are based on symetric linear phase FIR filters. These exhibit ringing that many consider not to sound good. There is no reason why you could not implement a digital filter that does not ring. You can implement a bessel filter in the digital domain [or you could use a minimum phase FIR filter].
Often when people create digital representations of analog filters, they use a zero order hold approximation [as Werner alludes to]. The zero order hold essentially assumes the signal is the same from the time it was sampled until the time the next sample is taken..... i.e. the typical flat-top digital representation of a signal. A way to improve this approximation, as Werner says, is to increase the sampling frequency. In my old days of designing control systems (one of many hats), we used to say that the zero order hold was only a good approximation if the sample frequency was 10 times the frequencies of interest....
However, higher sampling frequency is not the only way to improve the approximation. You can use higher order holds to create digital approximations of analog filters.
.... sorry if the theory is going over your head [I don't know your background], but this argument of non-os vs. os DACS I find is frought with a lack of understanding of the issues at hand or based on reviews of products that are not ideally implemented.
For the sake of simplicity I was keeping the IM issue under my hat, trying to keep everything, initially, at a more abstract level. That's also why I do not mention perceived sound quality of these technologies, as implementations differ marketedly. The AN DACs have quite steep filtering beyond 20kHz, the 47Labs machines have nothing at all, or perhaps a first-order.
Zero-order hold: was referring to the output stage of your typical R2R DAC chip (1541, 63, ...). To my knowledge no commercial audio DAC chips exist that implement higher order hold. However, there used to be a slew of Harman/Kardon CD-players that used two DACs per channel to join the dots with generated ramps.
As for the ringing. Read Nyquist. The ringing is required. If you make a digital reconstruction filter that does not ring, it is wrong.
Again, I am talking from a strictly mathematical point of view, striving to recover the original band-limited, then sampled, waveform.
Zero-order hold: was referring to the output stage of your typical R2R DAC chip (1541, 63, ...). To my knowledge no commercial audio DAC chips exist that implement higher order hold. However, there used to be a slew of Harman/Kardon CD-players that used two DACs per channel to join the dots with generated ramps.
As for the ringing. Read Nyquist. The ringing is required. If you make a digital reconstruction filter that does not ring, it is wrong.
Again, I am talking from a strictly mathematical point of view, striving to recover the original band-limited, then sampled, waveform.
Ok, but that is the main reason for digital filters so a discussion without looking at IM is soon over.
1541 is not fully R2R, just the first 6 bits. The other 10 are emitter-area-scaled transistors, with adjusted timing for simultaneous transistion when Vout=0 .
And of course dacs don't do hold, but filters do (if you mean that dacs create 'square' outputs, you're right of course). SAA7220 does 2x hold and 2x interpolation. Most others use a similar combination of 0 and higher orders.
Pre-ringing is only found in symmetrical fir filters. Symmetrical filters have constant phase. If you let the constant -phase requirement go and make a non-symmetrical minimum phase fir filter, there will be no pre-ringing.
Remco
If we implement a digital representation of a bessel filter, the amount of ringing will be minimal. On the other hand, the bessel filter will not be as good at attenuating the sampling artifacts. I could say the same for a minimum phase FIR. However, if we oversample enough, I think we can reach an optimum trade off.
Alvaius
Alvaius
For the last time: Nyquist and Shannon leave us no choice re low-pass filter. Anything that is not Sinc (or a valid approximation of it, being linear-phase and ringing), leads to erroneous reconstruction. But you have to go back to the original papers for this, and AFAIK they are not freely available on the net.
If you have access to Matlab with its full DSP package I invite you to build a waveform, band-limit it, note the resultant waveform down, then sample it, and subsequently reconstruct it with whatever you like. Only one particular type of reconstructor will yield the original waveform, for all kinds of input.
If you have access to Matlab with its full DSP package I invite you to build a waveform, band-limit it, note the resultant waveform down, then sample it, and subsequently reconstruct it with whatever you like. Only one particular type of reconstructor will yield the original waveform, for all kinds of input.
Infact it is well know that the human hearing organs produce IM distortion. The detection of IM distortion is used by audiologists to screen potential hearing impaired patients. But you need input frequencies that are transmitted by the relevant parts of the ear. I doubt that the hearing organ produces IM distortion with input frequencies beyond 20KHz.
gr,
Thijs
The only filtering requirement is being 'far' down at Fs/2. 'Far' being about 70-80 dB.
Leaving Fs at 44kHz means creating a brick wall filter between 20 and 22.05 kHz. The Fourier equivalent of a brick wall (box function) response is a sinc. Thus, a sinc function is required with non oversampling.
However, oversampling creates room for a less steep filter since the 'Far' down point is further away. Transition band is broader. So the filter can be less steep than a sinc. Hence oversampled systems don't necessarily use sincs.
And yes, the Matlab 'signal' toolbox simulates this quite good.
Remco
Leaving Fs at 44kHz means creating a brick wall filter between 20 and 22.05 kHz. The Fourier equivalent of a brick wall (box function) response is a sinc. Thus, a sinc function is required with non oversampling.
However, oversampling creates room for a less steep filter since the 'Far' down point is further away. Transition band is broader. So the filter can be less steep than a sinc. Hence oversampled systems don't necessarily use sincs.
And yes, the Matlab 'signal' toolbox simulates this quite good.
Remco
No way.
Shannon proves that f(t), provided it does not contain any frequencies above W (omega), can be decomposed into the following infinite sum of orthogonal functions:
f(t) = sum_over_n of
(
Xn x sin(Pi(2Wt-n))/(Pi(2Wt-n)
)
with Xn = f(n/2W), i.e. a train of
samples taken instantaneously, and '=' meaing 'being exactly equal to'.
The formal proof of this is also the formal proof for the simple fact that Sinc is the required reconstructor.
However, since Sinc can not be realised, over time, and given everyday practicalities (like the trivial task of building valve-based PCM systems in 1948!), the requirement was relaxed to "a steep cutoff at fs/2".
But this doesn't make that solution correct. Only "correct enough" for a given application.
Which was telephony at around that time
Shannon proves that f(t), provided it does not contain any frequencies above W (omega), can be decomposed into the following infinite sum of orthogonal functions:
f(t) = sum_over_n of
(
Xn x sin(Pi(2Wt-n))/(Pi(2Wt-n)
)
with Xn = f(n/2W), i.e. a train of
samples taken instantaneously, and '=' meaing 'being exactly equal to'.
The formal proof of this is also the formal proof for the simple fact that Sinc is the required reconstructor.
However, since Sinc can not be realised, over time, and given everyday practicalities (like the trivial task of building valve-based PCM systems in 1948!), the requirement was relaxed to "a steep cutoff at fs/2".
But this doesn't make that solution correct. Only "correct enough" for a given application.
Which was telephony at around that time
Werner, please do go into Matlab, or SystemView (my personal favourite!), or Hypersignal, or anything else, and run a digital square wave, over sampled, into a linear phase FIR, a minimum phase FIR, a digital representation of a Bessel Filter, and a digital representation of a Chebyshev. Then please post the waveforms for us to view.
Let's assume a 1KHz and 10KHz square wave for interest, and say 32x oversampling of a 44.1KHz signal. You could set the reconstruction filters to 22kHz for interest, or set them higher to reduce in band phase issues.
The digital representation of the bessel will illustrate very little ringing in this implementation compared to the linear phase FIR and chebyshev. That is not to say it is more accurate. In fact, if you look at the RMS error, it may be worse, but it does not ring.
In fact all this discussion of ringing, mathematical perfection, etc. is mute without a discussion of speakers and ears. For every unmeasured audio "difference", there is always a real reason for the sound difference, i.e. add a delayed signal to an original signal and there is no harmonic distortion added, but it sounds like crap! I have a curiousity that the issue with ringing is how it effects the operation of my mechanical speakers. IM distortion aside, a nice rounded step may be easier for my speakers to deal with. I am not saying that is true or not, it is just a thought.
Alvaius
Let's assume a 1KHz and 10KHz square wave for interest, and say 32x oversampling of a 44.1KHz signal. You could set the reconstruction filters to 22kHz for interest, or set them higher to reduce in band phase issues.
The digital representation of the bessel will illustrate very little ringing in this implementation compared to the linear phase FIR and chebyshev. That is not to say it is more accurate. In fact, if you look at the RMS error, it may be worse, but it does not ring.
In fact all this discussion of ringing, mathematical perfection, etc. is mute without a discussion of speakers and ears. For every unmeasured audio "difference", there is always a real reason for the sound difference, i.e. add a delayed signal to an original signal and there is no harmonic distortion added, but it sounds like crap! I have a curiousity that the issue with ringing is how it effects the operation of my mechanical speakers. IM distortion aside, a nice rounded step may be easier for my speakers to deal with. I am not saying that is true or not, it is just a thought.
Alvaius
'does not contain any' is nice in mathematics.
In real world engineering terms it means 'is lower than 70-80 dB below max'.
Therefore any real world filter will be an approximation. Even a sinc filter will suffer from component tolerances, noise and other effects. The form of the residual effects is a tradeoff, and that's what engineering is all about.
Nevertheless, any approximation will be more accurate if there is a larger transition band.
Remco
In real world engineering terms it means 'is lower than 70-80 dB below max'.
Therefore any real world filter will be an approximation. Even a sinc filter will suffer from component tolerances, noise and other effects. The form of the residual effects is a tradeoff, and that's what engineering is all about.
Nevertheless, any approximation will be more accurate if there is a larger transition band.
Remco
ultranalog said:
Not quite.
Oversampling on its own does nothing, except cheating by seamingly enlarging the sampled frequency domain (blowing up the unit circle, if you want). This enlargement is a necessity to allow the use of a sampled-domain filter (read digital filter). This digital filter still has the requirement of cutting sharply at the original fs/2. And to be entirely exact, the sum transfer function of the digital filter and any (softer) analogue filter must equal the Sinc function.
The sum total still cuts at fs/2 (20kHz), and if it doesn't you're deeply into Wadia/Legato Link territory.
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.