the best transformer in the audio signal path is no transformer.
the best capacitor in the audio signal path is no capacitor.
you can't go with tubes, and leave out caps and trafos, therefore best tubes in audio signal is no tubes.
Trafos are nice in DAC i/vs, as they fight many problems at once
- Mixing differential current outputs
- Ground detatching of output stage
- Bandwidth limiting without active components - you don't need an opamps to eat that MHz noise. They may act as aliasing filter too.
- Voltage "gain"
the best capacitor in the audio signal path is no capacitor.
you can't go with tubes, and leave out caps and trafos, therefore best tubes in audio signal is no tubes.
Trafos are nice in DAC i/vs, as they fight many problems at once
- Mixing differential current outputs
- Ground detatching of output stage
- Bandwidth limiting without active components - you don't need an opamps to eat that MHz noise. They may act as aliasing filter too.
- Voltage "gain"
Hi,
I do not find the ESS Dac's very interesting sonically. We have compared the full range of AMR digital products against a number of offerings using the ESS Sabre Reference DAC and played a little.
I see little reason to use these chip's on grounds of sonic performance. Stereoplay in Germany recently reviewed the AMR DP-777 and rated it's sonic performance higher than that of a competitor using the ES9018 (and also tubes), not by a lot, but if the ESS is a patch on the TDA1541A it should have killed the DP-777 (the CD-77 does)...
Given their extreme price premium over (say) AD, AKM, BB/TI, Cirrus and Wolfson's best their subjective performance is surprisingly little if any above the best of the rest. So don't expect much from my end. Though the attached ueber-hype may very well force a commercial choice to use them anyway or we may be asked to use it for OEM projects.
I am sure it is possible to do better than the implementations we have heard so far, but the same is true for pretty much any of the other manufacturers DAC Chip's as well.
I'm hoping on the Arda Tech DAC being a TDA1541 killer, we will see.
Ciao T
Looking forward to seeing your implementation of an ESS DAC at some point in the future.
I do not find the ESS Dac's very interesting sonically. We have compared the full range of AMR digital products against a number of offerings using the ESS Sabre Reference DAC and played a little.
I see little reason to use these chip's on grounds of sonic performance. Stereoplay in Germany recently reviewed the AMR DP-777 and rated it's sonic performance higher than that of a competitor using the ES9018 (and also tubes), not by a lot, but if the ESS is a patch on the TDA1541A it should have killed the DP-777 (the CD-77 does)...
Given their extreme price premium over (say) AD, AKM, BB/TI, Cirrus and Wolfson's best their subjective performance is surprisingly little if any above the best of the rest. So don't expect much from my end. Though the attached ueber-hype may very well force a commercial choice to use them anyway or we may be asked to use it for OEM projects.
I am sure it is possible to do better than the implementations we have heard so far, but the same is true for pretty much any of the other manufacturers DAC Chip's as well.
I'm hoping on the Arda Tech DAC being a TDA1541 killer, we will see.
Ciao T
I have yet to see a transformer that will apply the proper aliasing filtering needed - even for a 4x OS you need at least a 3-5th order filter. For NOS is impossible even with real analog filters to get a proper filtering.
http://home.mira.net/~gnb/mac-cdis/cd8.html
PS: That's why I like the Denon Alpha processing - it maintaines the original resolution after applying the OS. Some people (without proper math skills) asked what's the use of 20 bit when the initial signal is 16 bit. Truth is that every time you double the sampling frequency, you need to add one bit of digital resolution (at minimum, for simple linear interpolation), in order to keep the original samples intact. Otherwise you will loose the original resolution. 4x needs 18 bit, 8x needs 19 bit, 16x needs 20 bit to represent faithfully the original 1x 16 bit signal. Better interpolation algorithms need more bits per each OS step.
SAA7220 does 4x OS by truncation to 16 bit and therefore looses 2 bit of original information. The TDA1541 will output 14 bit from original signal and 2 bits of noise shaping. Some people are perfectly happy with that - I am not.
Last edited:
Given there are so many ways to mess up a 1541-based DAC and that most people still make suboptimal implementations (even though it's been around for ages), I think it'll take a while until he hear what the Sabre can really do.
For instance, only recently have they realized it actually sounds better in synchronous mode.
Hi,
I have yet to see a digital filter that does either. So what?
You are correct, you need to resolve the full 20 Bit and have the appropriately low jitter if you use an oversampling filter of this type of design to get at least MOST of your original 16 Bit information back. Note, the process can only be lossy (which anyone with the proper math skills knows), so the 20 Bit output does not contain all the original 16 bit information and you can not take the 20 Bit output and apply the same math in reverse to get your original 16 Bit input back.
First, there is no current digital filter that "keeps the original samples intact". Adding an extra bit resolution does not "keep the original samples intact" (which anyone with the proper math skills knows).
Second, you can implement digital filters in many ways. You do not need to have a greater output word length.
No, you only need 16 bit's for a 16 bit source, UNLESS you decide to implement a digital filter that interpolates between samples. Then you MUST have a DAC that is capable of fully resolving the new data to avoid loosing any more information than you have already lost.
However, there is no need to use interpolation between samples to apply digital filtering.
Based on the datasheet and application information I have for the SAA7220 it's digital filter neither employs wordlength expansion nor any subsequent truncation.
Of course Philips may have hidden such information from the engineering community and presented to do something else. Do you have information that substantiates such a deception on the part of Philips?
The TDA1541 will output the 16 Bit Signal from the digital filter. The filter in the SAA7220 is not equipped with any noise shaping algorithms any more than it truncates any data or uses word length expansion during the filtering.
All this is quite clear if one actually reads the Datasheet.
If we wanted, we could use a digital filter with a suitable algorithm to take an 18Bit wordlength at 1fs and output it to the TDA1541A in a way that produced a true 18 Bit output (Philips used this with the first generation TDA1540 DAC's and complementary filter).
You should note that these facts (using 4 * OS allows us to produce 16 Bit output using a 14 Bit DAC, not we require 18 Bit output from the filter and an 18 Bit DAC to produce oversampled 16 Bit output) are the precise reversal of your claims and that this has been public knowledge for over three decades.
Ciao T
I have yet to see a digital filter that does either. So what?
You are correct, you need to resolve the full 20 Bit and have the appropriately low jitter if you use an oversampling filter of this type of design to get at least MOST of your original 16 Bit information back. Note, the process can only be lossy (which anyone with the proper math skills knows), so the 20 Bit output does not contain all the original 16 bit information and you can not take the 20 Bit output and apply the same math in reverse to get your original 16 Bit input back.
First, there is no current digital filter that "keeps the original samples intact". Adding an extra bit resolution does not "keep the original samples intact" (which anyone with the proper math skills knows).
Second, you can implement digital filters in many ways. You do not need to have a greater output word length.
No, you only need 16 bit's for a 16 bit source, UNLESS you decide to implement a digital filter that interpolates between samples. Then you MUST have a DAC that is capable of fully resolving the new data to avoid loosing any more information than you have already lost.
However, there is no need to use interpolation between samples to apply digital filtering.
Based on the datasheet and application information I have for the SAA7220 it's digital filter neither employs wordlength expansion nor any subsequent truncation.
Of course Philips may have hidden such information from the engineering community and presented to do something else. Do you have information that substantiates such a deception on the part of Philips?
The TDA1541 will output the 16 Bit Signal from the digital filter. The filter in the SAA7220 is not equipped with any noise shaping algorithms any more than it truncates any data or uses word length expansion during the filtering.
All this is quite clear if one actually reads the Datasheet.
If we wanted, we could use a digital filter with a suitable algorithm to take an 18Bit wordlength at 1fs and output it to the TDA1541A in a way that produced a true 18 Bit output (Philips used this with the first generation TDA1540 DAC's and complementary filter).
You should note that these facts (using 4 * OS allows us to produce 16 Bit output using a 14 Bit DAC, not we require 18 Bit output from the filter and an 18 Bit DAC to produce oversampled 16 Bit output) are the precise reversal of your claims and that this has been public knowledge for over three decades.
Ciao T
Last edited:
^@ s3tup - NOS will add garbage noises exactly because there are no proper filters for that. Some like that garbage an call it "open sound"
Yeah, as well as paper-in-oil distortion as "sweet an mellow" sound, and opamp's distortion/instability as "dynamic and detailed" sound.
DACs are boring. Boring to death, as they are what they are - they produce what was originally recorded.
They should have deep well-defined soundstage in records which have it, have many pleasant details (which actually belong to recording, not the echoes of distortion), and have flat tonal ballance.
Last edited:
Hi,
NOTE, digital filters do not remove images!
They merely shift them to higher frequencies, where they tax the analogue circuitry more than if they where lower in frequency, if normal feedback circuitry of the "integrator" type is used.
Usually the frequency responses in the filter datasheets are cut off long before the return of the images becomes apparent, but that does not mean it's not there...
Ciao T
^@ s3tup - NOS will add garbage noises exactly because there are no proper filters for that. Some like that garbage an call it "open sound"
NOTE, digital filters do not remove images!
They merely shift them to higher frequencies, where they tax the analogue circuitry more than if they where lower in frequency, if normal feedback circuitry of the "integrator" type is used.
Usually the frequency responses in the filter datasheets are cut off long before the return of the images becomes apparent, but that does not mean it's not there...
Ciao T
I'm still looking for a DAC that will produce the recorded music as it is, without any additions and omissions…
I forgot to ask how some can Over Sample 1 bit (LSB) variation in signal, without interopolation and increasing the resolution? If resolution remains identical, the new OS value will the same as original (equal either with the first or with the second sample), therefore the final signal will consist in two identical samples one after another.
That's why, in order to maintain the original samples, the new, interpoated, sample needs to be interpolated between the original samples and requires increased bit depth.
Sure, it can be done without interpolation, but the end result will be sub-optimal at LSB levels, therfore loosing resolution. That's how SAA7220 does it.
@s3tup - True. That's why you need better resolution that the current 16 recordings and that's why I like the SACD recordings. But that is off topic, don't want to upset people.
That's why, in order to maintain the original samples, the new, interpoated, sample needs to be interpolated between the original samples and requires increased bit depth.
Sure, it can be done without interpolation, but the end result will be sub-optimal at LSB levels, therfore loosing resolution. That's how SAA7220 does it.
@s3tup - True. That's why you need better resolution that the current 16 recordings and that's why I like the SACD recordings. But that is off topic, don't want to upset people.
If you take OS x2, pass 2 samples with 1 LSB difference, and then low-pass it, how many bits you get from 2 bits? 3. you get a 0, 0.5, and 1 bits.
OTOH, if you pass a digital signal thru IIR/FIR, you loose bits in precision during multiplication.
It's up to specific DF to guestimate the gains and losses of bits precision.
OTOH, if you pass a digital signal thru IIR/FIR, you loose bits in precision during multiplication.
It's up to specific DF to guestimate the gains and losses of bits precision.
Hi,
Digital filtering works not by attempting to ïnterpolate "between individual samples"; as in sample 1 is 1LSB and sample 2 is 2LSB, so for 4 * OS I need samples at 1LSB, 1.25LSB, 1.5LSB and 1.75LSB and then we have the 2LSB sample.
It instead stores many samples (in the case of the SAA7220 30 samples) and applies a set of "weights" to each sample. In the case of the SAA7220 this is done twice but in opposite directions - resulting in the familiar symmetrical impulse response. The "weighted" samples then summed into a new value which is output as output sample.
The Philips datasheet is actually quite clear and lucid about how their chip works.
That depends on the chosen window and filter function and signal. As we never only look at two samples, but in the case of the rather pedestrian SAA7220 at 30 Samples the case you describe is extremely unlikely (I cannot be bothered to calculate how unlikely).
It should be noted that the SAA7220 holds by far fewer samples than modern filters and more samples contribute to each output samples value in modern designs than do in this old part.
Let me repeat this, to be clear.
The original samples are not retained by most practical digital filters, in fact I know of non that does.
If you have information that shows a practical filter that does preserve the original samples, please provide this information with actual positive proof.
So, now suddenly the SAA7220 no longer uses truncation and noise shaping?
So now it does suddenly, somehow no longer interpolate instead?
How quickly you change your tune.
But your claims are still severely non-congruent with how things actually work. As said, the Philips Datasheet is quite lucid, it may be worth reading it.
Well, alas, it does interpolate, where the filter function requires it. If it needs a value that for one input sample is less than one LSB it simply uses the time domain to represent intermediate values. Again, this is the way digital filters work.
There is no magic.
And no, the SAA7220 does NOT loose resolution, as it designed to work with a 16 Bit DAC.
That is so ironic.
SACD or DSD is by far worse for adding high ultrasonic rubbish than 16 Bit Non-Os.
I normally find myself on opposite sides of arguments with Mr. Lipshitz, but the Lipshitz/Vanderkooy paper(s) on DSD really should be mandatory reading for understanding the tradeoffs and issues in this particular design and for Delta-Sigma parts overall.
I once wrote a Spreadsheet to demonstrate the effect of Zero Order Hold (that is non-oversampling), primarily as an excercise in shutting up those ejits who promptly tried to sow FUD by claiming "Non-Oversampling must blow speakers so do not even try it").
I'd post it here, but it is 10MB file.
It is instructive to compare the output from that to the output of a DSD modulator, especially if the ZOH system is fitted with a 50KHz lowpass.
For totally unfiltered 44.1KHz ZOH DSD exceed the ultrasonic noise a DIGITAL FULL SCALE SIGNAL in the ZOH system with a 50KHz lowpass (and ZOH rolloff compensation) would produce at around 80KHz.
The Ultrasonic noise of DSD is (ideally) not affected by the signal (in reality it is of course) and this constant, always present regardless what the signal level is.
Music rarely contains anything that is full scale steady state (a crest factor of 10dB is usually observed even with the most heavily compressed music) and the ultrasonic images produced by ZOH have signal level depend levels.
So in practice, when playing music, ZOH add more ultrasonic noise than DSD only in the octave between 20KHz and 40Khz. And this noise in this band has some interesting properties of it's own that exceed the scope of this discussion.
Above around 40KHz (depending on the track played lower) DSD in fact adds more ultrasonic noise than ZOH PCM.
Ciao T
Digital filtering works not by attempting to ïnterpolate "between individual samples"; as in sample 1 is 1LSB and sample 2 is 2LSB, so for 4 * OS I need samples at 1LSB, 1.25LSB, 1.5LSB and 1.75LSB and then we have the 2LSB sample.
It instead stores many samples (in the case of the SAA7220 30 samples) and applies a set of "weights" to each sample. In the case of the SAA7220 this is done twice but in opposite directions - resulting in the familiar symmetrical impulse response. The "weighted" samples then summed into a new value which is output as output sample.
The Philips datasheet is actually quite clear and lucid about how their chip works.
That depends on the chosen window and filter function and signal. As we never only look at two samples, but in the case of the rather pedestrian SAA7220 at 30 Samples the case you describe is extremely unlikely (I cannot be bothered to calculate how unlikely).
It should be noted that the SAA7220 holds by far fewer samples than modern filters and more samples contribute to each output samples value in modern designs than do in this old part.
Let me repeat this, to be clear.
The original samples are not retained by most practical digital filters, in fact I know of non that does.
If you have information that shows a practical filter that does preserve the original samples, please provide this information with actual positive proof.
So, now suddenly the SAA7220 no longer uses truncation and noise shaping?
So now it does suddenly, somehow no longer interpolate instead?
How quickly you change your tune.
But your claims are still severely non-congruent with how things actually work. As said, the Philips Datasheet is quite lucid, it may be worth reading it.
Well, alas, it does interpolate, where the filter function requires it. If it needs a value that for one input sample is less than one LSB it simply uses the time domain to represent intermediate values. Again, this is the way digital filters work.
There is no magic.
And no, the SAA7220 does NOT loose resolution, as it designed to work with a 16 Bit DAC.
That is so ironic.
SACD or DSD is by far worse for adding high ultrasonic rubbish than 16 Bit Non-Os.
I normally find myself on opposite sides of arguments with Mr. Lipshitz, but the Lipshitz/Vanderkooy paper(s) on DSD really should be mandatory reading for understanding the tradeoffs and issues in this particular design and for Delta-Sigma parts overall.
I once wrote a Spreadsheet to demonstrate the effect of Zero Order Hold (that is non-oversampling), primarily as an excercise in shutting up those ejits who promptly tried to sow FUD by claiming "Non-Oversampling must blow speakers so do not even try it").
I'd post it here, but it is 10MB file.
It is instructive to compare the output from that to the output of a DSD modulator, especially if the ZOH system is fitted with a 50KHz lowpass.
For totally unfiltered 44.1KHz ZOH DSD exceed the ultrasonic noise a DIGITAL FULL SCALE SIGNAL in the ZOH system with a 50KHz lowpass (and ZOH rolloff compensation) would produce at around 80KHz.
The Ultrasonic noise of DSD is (ideally) not affected by the signal (in reality it is of course) and this constant, always present regardless what the signal level is.
Music rarely contains anything that is full scale steady state (a crest factor of 10dB is usually observed even with the most heavily compressed music) and the ultrasonic images produced by ZOH have signal level depend levels.
So in practice, when playing music, ZOH add more ultrasonic noise than DSD only in the octave between 20KHz and 40Khz. And this noise in this band has some interesting properties of it's own that exceed the scope of this discussion.
Above around 40KHz (depending on the track played lower) DSD in fact adds more ultrasonic noise than ZOH PCM.
Ciao T
Hi,
You really need to start with the microphones and ADC's for that.
BUT, just as a microphone is not directly a speaker in reverse and just as it is not directly an electrical analog of the ear, so is the DAC not an ADC in reverse nor is the signal leaving the speaker a direct replica of the signal that arrived at the microphone...
What is perhaps most amazing in all this "not a precise replica" is that the end of it, when sitting down, not only are we able to hear by far more than bare snatches of a tune through noise and distortion, we can even hear a great amount of detail, all the way to positioning of instruments...
If I showed anyone the actual signal going to the two speakers on a 'scope and asked anyone which part of that line is the first violin, which part is the second and which is the trombone, everyone would just shrug and tell me to shut up.
If I asked the AP2 designers to give me a test that could tell me which part of that wriggly line is Violin and which trombone they would look at me with more astonishment than they would look at ET if he stepped right in front of them.
Yet I can apply this wriggely line to a very imperfect speaker and voila, there is the first violin, there is the second, there is the cellist, there is the trombone and so on... I can even often pretty well point at the player that was a bit off on note or timing.
So perhaps we need by far more astonished that we hear anything resembling music, rather than critical that we do not get "the music as it is". How this bunch or wires, vaccuum or silicone and other stuff managed to pull off this amazing feat has been obsessing and foxing me ever since I found out at age three or so that there where no little musicians and announcers in the Radio (by the time we got TV I had understood the nature of it as artifice, if not in detail how it worked).
Ciao T
You really need to start with the microphones and ADC's for that.
BUT, just as a microphone is not directly a speaker in reverse and just as it is not directly an electrical analog of the ear, so is the DAC not an ADC in reverse nor is the signal leaving the speaker a direct replica of the signal that arrived at the microphone...
What is perhaps most amazing in all this "not a precise replica" is that the end of it, when sitting down, not only are we able to hear by far more than bare snatches of a tune through noise and distortion, we can even hear a great amount of detail, all the way to positioning of instruments...
If I showed anyone the actual signal going to the two speakers on a 'scope and asked anyone which part of that line is the first violin, which part is the second and which is the trombone, everyone would just shrug and tell me to shut up.
If I asked the AP2 designers to give me a test that could tell me which part of that wriggly line is Violin and which trombone they would look at me with more astonishment than they would look at ET if he stepped right in front of them.
Yet I can apply this wriggely line to a very imperfect speaker and voila, there is the first violin, there is the second, there is the cellist, there is the trombone and so on... I can even often pretty well point at the player that was a bit off on note or timing.
So perhaps we need by far more astonished that we hear anything resembling music, rather than critical that we do not get "the music as it is". How this bunch or wires, vaccuum or silicone and other stuff managed to pull off this amazing feat has been obsessing and foxing me ever since I found out at age three or so that there where no little musicians and announcers in the Radio (by the time we got TV I had understood the nature of it as artifice, if not in detail how it worked).
Ciao T
Wow, that is the best exposition of an idea I have had for a long time. Especially considering so many in high end say the ear is the best instrument, when, in fact, it appears to be the worst. I think that is the root of the problem, we don't want an exact replica of what was recorded, we want a replica of what was performed. Luckily, the brain is easily fooled, and things like sound-staging, imaging and depth, exist because our brains take shortcuts. But, because of the shortcuts, our brain-sense experience is easily fooled. I think it would be nice to learn to measure what is important the brain-hearing sense, because until we do, there will continue to be arguments about un-measurable distortion sounding worse or better than se 300b amps, etc. Even so, I am going to attempt to make a pcb for the DAC you describe here, I will probably start a new thread considering I know very little about digital pcb layout.
Hardly likely - the technical info they're so far put out shows it as being based on a resistor ladder. R-2R anyone?
I think this must be false for the simple reason that it adds an offset. So a full scale input would be sure to clip the output were no resolution lost. Of course the loss of resolution is small, but its very definitely there.
<edit> A second and more serious loss of resolution comes about because it has a non-flat frequency response, compensating as it does for the downstream filter.
(the rest of your post I by and large agree with and +1 for highlighting some of the BS about DSD).
Last edited:
Hi,
Read carefully, the DC offset is tied in with the DAC Design.
But yes, it will in theory introduce a 5% error somewhere. But only with signals at 0dBFS.
A LONG, long time ago in a massive meta-thread on DAC design I pointed out that 16 Bit Audio was not capable of representing a -96dBFS (or indeed anything below -90.31dBFS) unless using "illegal" means such as DC offset etc.
Adding such DC offset can be a very pragmatic way on working around this limitation...
Well, as it offers 4 * OS it should be able to represent 18 Bits using the 16 Bit DAC, so for 16 Bit inputs it SHOULD not loose any resolution by adding 5% offset and a small amount of EQ with suitable scaling.
BTW,the scaling, EQ and DC offset are "designed in" into the coefficients of the filter, before anyone gets ideas on how to mis-read and mis-interpret the DC offset and Frequency response shaping.
But I agree, these implementation details are questionable and I would not use the SAA7220 in any new design. There are some options for digital filters that are quite okay, though with CD standard and similar (48KHz) recordings I almost always prefer ZOH with EQ over Digital Filtering, no matter what kind of filter.
Ciao T
Read carefully, the DC offset is tied in with the DAC Design.
But yes, it will in theory introduce a 5% error somewhere. But only with signals at 0dBFS.
A LONG, long time ago in a massive meta-thread on DAC design I pointed out that 16 Bit Audio was not capable of representing a -96dBFS (or indeed anything below -90.31dBFS) unless using "illegal" means such as DC offset etc.
Adding such DC offset can be a very pragmatic way on working around this limitation...
Well, as it offers 4 * OS it should be able to represent 18 Bits using the 16 Bit DAC, so for 16 Bit inputs it SHOULD not loose any resolution by adding 5% offset and a small amount of EQ with suitable scaling.
BTW,the scaling, EQ and DC offset are "designed in" into the coefficients of the filter, before anyone gets ideas on how to mis-read and mis-interpret the DC offset and Frequency response shaping.
But I agree, these implementation details are questionable and I would not use the SAA7220 in any new design. There are some options for digital filters that are quite okay, though with CD standard and similar (48KHz) recordings I almost always prefer ZOH with EQ over Digital Filtering, no matter what kind of filter.
Ciao T
I can't yet see how that's relevant here.
No, the offset is 0x0020 whereas the DAC's fullscale is 0x7FFF - that reduces the maximum output by 1>>10 or 0.1%.
I don't buy it - 'SHOULD' != 'is'
There's no way an offset could be designed in (with or without scare quotes) to the coefficients, since those are multipliers and offset requires merely the addition of a constant. But yes, the 12bit coefficients contain enough to effect the EQ - what's not clear to me is the size of the accumulator. 16bits*12bits gives a 27bit result, and 30 such results accumulated could in theory extend to 32bits.
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.