Sometime this is just the layout and not the best & bigger efforts maid on the technologie alone which is the weak link. Bad main caps quality decoupling e.g, or little difference with a resistor (plain smd resistor are sometimes prefered than low noise multilayer ones like the Sussumu RR in signal path, etc...)
If NPOs doesn't have too much problems, higher values with X7R like 1 Uf and more are poor choice but the paper theory (better inductance but often a bigger one with a choosed dielectric or in another construction give a better subjective sounding result. experience shows that this difference can go to a bad sounding to a good one if the layout is good at the beginning ! It's not magic, it's emipirical after tests; my modest experience showed me than e.g. two polymers of the same pitch leads and same ESR...give a total subjective different listening result ! : intereactions with passive between each others (caps and resistors parts are more difficult than the theory would say...and scope shows ! IHMO)
An exemple can be given with such old technolies which marry new technolgies as FGPAs, etc and old one like Abrax says (TDA1541 e.g.) and others like tubes can give also the best soundings results : I'm sure than a AMR CD-77 vs the little brand new Chord could be an interresting experience.
And what to say of some too simple choice about PS sometimes !
Here the SOTA development and the open mind ot the OP (I mean Soekris as example in the other thread) could allow good final result, after some different final batchs with differents passive parts and programmed filters... because not sure e.g. an overesampled (by the device) red book files can profit on a such low noise floor ! Some others subjectives qualities like Abrax says are important as he worked a lot e.g. on the importance of the power supply... : dynamics, etc...
Of course it's just a two cents and humble opinion. Of course SOTA enginneering and Layout is important... but I notice often the scope measurement is first prefered to reduce the listening tests because time & money is often the main factor.
Devices like the AMR CD-77 and long experience as man like T. Loesch have show it shouldn't be the way !
Hé ABrax why not spend 17 AUD to get the Distinction-1541 board here and find a nice TDA1541 in a second hand device at 15 USD ?
Sorry to be OT...
If NPOs doesn't have too much problems, higher values with X7R like 1 Uf and more are poor choice but the paper theory (better inductance but often a bigger one with a choosed dielectric or in another construction give a better subjective sounding result. experience shows that this difference can go to a bad sounding to a good one if the layout is good at the beginning ! It's not magic, it's emipirical after tests; my modest experience showed me than e.g. two polymers of the same pitch leads and same ESR...give a total subjective different listening result ! : intereactions with passive between each others (caps and resistors parts are more difficult than the theory would say...and scope shows ! IHMO)
An exemple can be given with such old technolies which marry new technolgies as FGPAs, etc and old one like Abrax says (TDA1541 e.g.) and others like tubes can give also the best soundings results : I'm sure than a AMR CD-77 vs the little brand new Chord could be an interresting experience.
And what to say of some too simple choice about PS sometimes !
Here the SOTA development and the open mind ot the OP (I mean Soekris as example in the other thread) could allow good final result, after some different final batchs with differents passive parts and programmed filters... because not sure e.g. an overesampled (by the device) red book files can profit on a such low noise floor ! Some others subjectives qualities like Abrax says are important as he worked a lot e.g. on the importance of the power supply... : dynamics, etc...
Of course it's just a two cents and humble opinion. Of course SOTA enginneering and Layout is important... but I notice often the scope measurement is first prefered to reduce the listening tests because time & money is often the main factor.
Devices like the AMR CD-77 and long experience as man like T. Loesch have show it shouldn't be the way !
Hé ABrax why not spend 17 AUD to get the Distinction-1541 board here and find a nice TDA1541 in a second hand device at 15 USD ?
Sorry to be OT...
Last edited:
A null state should be defined by around 0 dB SPL, right? See D. Mills post here.
Thus, without dither, we are moving from the "digital black" as you called it of 0 dB SPL to the maximum SPL the bit depth is capable of and thereby define the dynamic range.
I'm pretty sure audio bit depth should be defined by it's dynamic range without dither.
Introducing dither is a third parameter, which introduces a constant noise floor, which is higher than the digital black.
Let's just nevermind the 64-bit theory.
- When we reach a true 24-bit or higher ADC, we can stop applying dither, since at that point it has zero function.
- We can stop applying dither at 21-bit ADC as well, if we prefer the sound without it.
- We can stop applying dither to DSD recordings as well, if we prefer the sound without it.
- Audio bit depth is dynamic range, from digital silence to the loudest level.
- Audio bit depth + dither, is defined via the shaped noise floor, for example 120 dB dynamic range for 16-bit, according to Wikipedia.
- 4-bit / 44.1 kHz would be perfect for high-end audio if it had no quantization noise and we didn't care very much about dynamic range.
There, I have scientifically conluded the above 6 points from this thread.
I am really not sure about the sixth. Feel free to mathematically invalidate any of them if you can / wish to.
I may away from DIYaudio for a few days now.
See you!
Last edited:
I'm uncertain of the point you are attempting to make. My only point about an quantizer's null state is that it makes for an false reference point in determining dynamic range. The low end of the dynamic range lies at level of the smallest controllable change from input to output. In the analog domain, the thermal noise floor essentially sets that minimum controllable level of change. In the digital domain, however, the quantization noise/error sets the lowest level of controllable or resolvable change, at least, it does so within a single sample. The key word is, controllable.
I before used the analogy of automobile engine noise. The dynamic range of an automobile engine's sound would be measured from full throttle down to idle, not from full throttle down to not running. You cannot control the output of an engine which is not running. Not an ideal analogy, I've no doubt. Perhaps, think about it this way. An 1-bit quantizer has an dynamic range of only about 6dB, which equates to an signal amplitude ratio of only 2:1. It would not be correct, however, to conceptualization this 1-bit quantizer as having some huge dynamic range simply because one of it's two states could be said to produce a null output level. That output level is resolvable to only two levels. Digital domain resolution and quantization noise essentially describe the same parameter.
Last edited:
Hi,
Sorry I don't follow you, if the 1-bit quantizer has a dynamic range between digital silence and 6 dB then the dynamic range is 6 dB is not it?
Sorry I don't follow you, if the 1-bit quantizer has a dynamic range between digital silence and 6 dB then the dynamic range is 6 dB is not it?
The dynamic range is 6dB between the full-scale level and the minimum controllable/resolvable level. The 'digital silence' is not resolvable beneath the full-scale signal level minus 6dB, so the dynamic range is 6dB and the quantization noise floor is at -6dB. Much of DSP science is non-intuitive.
Dynamic range is the ratio between the lowest and the highest level. If you define the lowest level as zero, that ratio is always infinite. Anything divided by zero is by definition infinite.
Jan
You know what, this thread really went through a maze, but I think I may have found the answer.
DSD has a 1-bit resolution and it can effectively avoid quantization noise via 64x oversampling.
I've seen at least one DAC with a -140 dB noise floor, I think "The Wire" amplifier here at DIYaudio has a similar noise floor if I'm not mistaken.
DSD, has a maximum achievable dynamic range of 120 dB, for some reason.
Human hearing, can perceive a slightly higher dynamic range than that, apparently.
Thus, the --highest perceivable resolution without quantization noise--, is 2-bit?
Since, we apparently need 2-bit to succeed 120 dB.
The highest SPL resolution, period, is from silence to 194 dB, according to this link, not sure if correct or not ----- Decibel (Loudness) Comparison Chart
When I started this thread, I didn't know audio bit-depth resolution was only SPL, I thought it interacted with time or THD as well, for instance as XXHighEnd states, which we've discussed.
Any comments? Not very clear is it.
DSD has a 1-bit resolution and it can effectively avoid quantization noise via 64x oversampling.
I've seen at least one DAC with a -140 dB noise floor, I think "The Wire" amplifier here at DIYaudio has a similar noise floor if I'm not mistaken.
DSD, has a maximum achievable dynamic range of 120 dB, for some reason.
Human hearing, can perceive a slightly higher dynamic range than that, apparently.
Thus, the --highest perceivable resolution without quantization noise--, is 2-bit?
Since, we apparently need 2-bit to succeed 120 dB.
The highest SPL resolution, period, is from silence to 194 dB, according to this link, not sure if correct or not ----- Decibel (Loudness) Comparison Chart
When I started this thread, I didn't know audio bit-depth resolution was only SPL, I thought it interacted with time or THD as well, for instance as XXHighEnd states, which we've discussed.
Any comments? Not very clear is it.
Last edited:
Hello Kastor L
I don't know if we can call it TIM distortion, simply it looks like the TIM distortion of analog audio circuits. I say that because there is a difference between a TIM distorted analog square wave with this taken from the output of the oversampling (or interpolating) DAC. Particularly in analog circuits the overshoot occurs at the rising edge of square waveform and is followed by ringings with reduced amplitude - down to zero - up to the end of duty cycle. While in DAC the ringings are minimized at the middle of duty cycle and then are growing up again up to the falling edge of waveform.
Last edited:
Hmm... Looking in fast at the thread of this Danish who made a NOS DAC, i can say that the conception of this idea is not so difficult, while its implementation in actual hardware is amazingly complex. We did a similar process in the past in light dimmers managed through the DMX512 communication protocol. We used a 8bit DAC0808 managed by a micro to decode the analog level at the receiving device, or inverserly a ladder network combined with a parallel in / serial out shift register to encode the analog level into the transmiting device. What needs a NOS DAC is a Manchester decoder at the input (SPDIF signal stream is Manchester encoded i.e. each bit is biphasic) to recover the 3 PCM data lines: bit clock, word clock (L-R clock) and data. A microcontroller (the Danish uses a FPGA device to this) could collect all necessary information of the signal - just from the two clocks - and accordingly it can manage a set of serial in / parallel out shift registers which off the serial in will be the PCM data line while resistors connected across the parallel outputs and GND will convert the audio data bit stream to equivalent voltage level. Finally a summing network will add all momentary voltage levels to reconstruct the original audio signal.
I know that is some crude this explanation but i think that is the general concept of a NOS DAC.
Firstly, there's a "free" DSP book here (free except for one's time and effort to read and understand it):
The Scientist and Engineer's Guide to Digital Signal Processing
I'll respond to some other things, even though perhaps ...
Human hearing, from the least discernable sound to the threshold of pain, has a dynamic range of 120dB. 20 bits * 6dB = 120dB, so 20 bits is enough to cover the range of human hearing.
If you don't put an anti-aliasing filter before the ADC, any ultrasonic energy in the signal is DEFINITELY and ABSOLUTELY "reflected back" into the audible range by the ADC.
If you don't put a reconstruction filter on the output of the DAC (as apparently shown in that sine wave you posted), in an ideal world it won't make any difference, but in a practical world every amplifier has distortion, and usually the higher frequency (such as ultrasonics) the more distortion, and the ultrasonic signal can be intermodulated with any other signal present and these intermodulation products can be audible.
With dither (the correct amount of noise added before truncating to 4 bits), the note will fade without a significant change in tone from what a much better reproduction of a piano note would, and it will fade into the noise, though it will be audible well below the level where the truncated note disappeared.
The Scientist and Engineer's Guide to Digital Signal Processing
I'll respond to some other things, even though perhaps ...
Wait a second, can we get a technical definition of blockiness? Otherwise this 100-bit figure is meaningless.
I ESPECIALLY want to know how to filter out IMD. I'll apply for the patent...
I'll try a gross approximation. Grossly speaking, every added bit of resolution gives the capability to reproduce a tone 6dB lower than the previous resolution. Eight bits goes down to -48dB, 16 bits goes down to -96dB.
Human hearing, from the least discernable sound to the threshold of pain, has a dynamic range of 120dB. 20 bits * 6dB = 120dB, so 20 bits is enough to cover the range of human hearing.
It depends.
If you don't put an anti-aliasing filter before the ADC, any ultrasonic energy in the signal is DEFINITELY and ABSOLUTELY "reflected back" into the audible range by the ADC.
If you don't put a reconstruction filter on the output of the DAC (as apparently shown in that sine wave you posted), in an ideal world it won't make any difference, but in a practical world every amplifier has distortion, and usually the higher frequency (such as ultrasonics) the more distortion, and the ultrasonic signal can be intermodulated with any other signal present and these intermodulation products can be audible.
An interesting test would be take a piano note or chord that gets held for 30 seconds or a minute, so it decays to well under the -24dB of the smallest 4-bit reproduction. With truncation (the buzzword for not using dither), while the note fades it will have more and more distortion until it is finally something like a square wave, and then suddenly disappear.
With dither (the correct amount of noise added before truncating to 4 bits), the note will fade without a significant change in tone from what a much better reproduction of a piano note would, and it will fade into the noise, though it will be audible well below the level where the truncated note disappeared.
Or when the music passage drops below the level that moves any of the four most-significant bits (which could be as high as roughly, grossly -24dB), the output becomes "truly silent."
The waveform produced by xx-bit quantized undithered music, either recorded or synthetically made in a computer I suppose, played on a 16-bit or 24-bit R2R DAC, without a reconstruction filter.
It seems like 21-bit is satisfactory in that area if we are considering the level at which self-dithering is present and quantization noise is very low.
The reason why a 24-bit R2R DAC apparently sounds better with upsampling or oversampling is still a concern.
I'm assuming at this moment it pushes away the IMD so it's not reflected back, or it improves the curvature of a sine.
A cancellation chip which is sensitive to IMD specifically and sends that content to a different amplifier?
In that case upsampling a Nos DAC without a reconstruction filter can reduce the THD.
Which is what XXHighEnd says and most likely what I am listening to when my Nos DAC sounds better 2x upsampled via software.
When I was playing with the 4-bit file I couldn't hear any difference but with normal classical music for instance it does sound a little nicer 2x upsampled.
XXHighEnd says when he upsamples 16x then the sine waves are pushed apart let's say from 10 Hertz to 700 kHz so then there are "less sines" available in the 20-20 area.
I'll send the link in my next post, then you can hear that the vocals are intact but the string instrument is covered up by the quantization noise.
If I understand it correctly the string instrument is still there, intact, but the quantization noise is too loud?
Yes with truncated 10-bit music, with a dynamic range of 60 dB then the musical content will shift from quiet to loud at a maximum of 60 dB, plus the digital silence.
In that sense, during the silent passages then the media file can be raised to a volume of 110 dB without hearing any noise!
Actually, I think the digital silence if you do not want to consider it as zero, includes the thermal / flicker noise from the ADC? Which we can say in this example is -120 dB.
Last edited:
Here is where I found the 4-bit file......
http://www.tnt-audio.com/sorgenti/dither_e.html
Here it is with zero dither, dither and shaped dither......
http://www.audiochrome.net/clips/Venice_4b_dither.mp3
http://www.audiochrome.net/clips/Venice_4b_nodither.mp3
http://www.audiochrome.net/clips/Venice_4b_noiseshapeE.mp3
What is really needed is a 16-bit version of that file as well.
Nonetheless what this test really taught me is how normal 4-bit / 44.1 kHz audio sounds.
The bit-depth doesn't really seem to have very much impact, only the 44,100 samples per second.
It's the 44,100 samples per second sound here which I and I think the vast majority of listeners are referring to when we say "resolution".
The purity of a 1 kHz sine is another form of resolution, which is calculated via harmonic distortion.
The reconstruction filter will vary in resolution as well.
Here is an article by technical staff, concerning their DAC with 7 kinds of reconstruction filters ----- http://www.audiostream.com/content/what-are-digital-filters-and-why-are-they-requried-todays-audio-dacs-resonessence-labs-tech-3
Here is another one ----- http://www.diyaudio.com/forums/digital-line-level/259397-new-chord-hugo-dac-2.html
"if you took the view that 16 bit sinc function coefficients were OK, thus ensuring time domain errors were below 16 bit, then you would be looking at getting on for 1,000,000 taps for an 8 times OS filter!"
______
"High resolution" media is usually referring to 24-bit / 192 kHz or higher.
16-bit / 192 kHz media would be called high resolution as well, hence speaking of the "increased resolution" in the higher hypersonic frequencies.
According to Wikipedia, DSD audio is running at 1⁄32768th of 16-bit audio resolution!
In the same article, they write
"A DSD-CD however does not achieve the same sound resolution as SACD because the high-sample rate, low-resolution DSD sound has to be converted to 44.1 kHz, 16-bit PCM in order to be compliant with the Red Book audio CD standard. DSD-CDs are fully compatible with CD."
O.k., so first SACD audio is running at 32,768 times less resolution than a CD, then in the next moment, if we try to convert a DSD to a CD, suddenly it "can not achieve the same resolution as an SACD".
......
Audio bit depth resolution is volume / SPL resolution, insofar as it does not interact with a different parameter.
Just to repeat, then there is
- sample-per-second resolution
- THD / IMD resolution
- hypersonic frequency resolution
- reconstruction filter resolution
Then there is the Nos DAC which sounds "more natural" which I assume is "higher resolution" as well since the natural sound of reality is in fact the highest
resolution.
Thus
- Non-interpolating resolution
Perhaps......
- R2R DAC resolution
In summary what is the highest resolution without quantization noise?
It seems like resolution needs to be strictly defined, first.
http://www.tnt-audio.com/sorgenti/dither_e.html
Here it is with zero dither, dither and shaped dither......
http://www.audiochrome.net/clips/Venice_4b_dither.mp3
http://www.audiochrome.net/clips/Venice_4b_nodither.mp3
http://www.audiochrome.net/clips/Venice_4b_noiseshapeE.mp3
What is really needed is a 16-bit version of that file as well.
Nonetheless what this test really taught me is how normal 4-bit / 44.1 kHz audio sounds.
The bit-depth doesn't really seem to have very much impact, only the 44,100 samples per second.
It's the 44,100 samples per second sound here which I and I think the vast majority of listeners are referring to when we say "resolution".
The purity of a 1 kHz sine is another form of resolution, which is calculated via harmonic distortion.
The reconstruction filter will vary in resolution as well.
Here is an article by technical staff, concerning their DAC with 7 kinds of reconstruction filters ----- http://www.audiostream.com/content/what-are-digital-filters-and-why-are-they-requried-todays-audio-dacs-resonessence-labs-tech-3
Here is another one ----- http://www.diyaudio.com/forums/digital-line-level/259397-new-chord-hugo-dac-2.html
"if you took the view that 16 bit sinc function coefficients were OK, thus ensuring time domain errors were below 16 bit, then you would be looking at getting on for 1,000,000 taps for an 8 times OS filter!"
______
"High resolution" media is usually referring to 24-bit / 192 kHz or higher.
16-bit / 192 kHz media would be called high resolution as well, hence speaking of the "increased resolution" in the higher hypersonic frequencies.
According to Wikipedia, DSD audio is running at 1⁄32768th of 16-bit audio resolution!
In the same article, they write
"A DSD-CD however does not achieve the same sound resolution as SACD because the high-sample rate, low-resolution DSD sound has to be converted to 44.1 kHz, 16-bit PCM in order to be compliant with the Red Book audio CD standard. DSD-CDs are fully compatible with CD."
O.k., so first SACD audio is running at 32,768 times less resolution than a CD, then in the next moment, if we try to convert a DSD to a CD, suddenly it "can not achieve the same resolution as an SACD".
......
Audio bit depth resolution is volume / SPL resolution, insofar as it does not interact with a different parameter.
Just to repeat, then there is
- sample-per-second resolution
- THD / IMD resolution
- hypersonic frequency resolution
- reconstruction filter resolution
Then there is the Nos DAC which sounds "more natural" which I assume is "higher resolution" as well since the natural sound of reality is in fact the highest
resolution.
Thus
- Non-interpolating resolution
Perhaps......
- R2R DAC resolution
In summary what is the highest resolution without quantization noise?
It seems like resolution needs to be strictly defined, first.
Last edited:
Technically speaking the Danish DAC is a 24-bit / 28-bit R2R network I think.
Have a look at the Ring DAC ----- Ring DAC | dCS
Clearly this vintage technology is still at the highest levels of audio, I do not necessarily support it, I'm just looking at the reasons for why.
I don't think the reason is interpolation, since even the Nos supporters tend to think that sounds fine.
It seems more likely the reason is R2R.
In some audio places, the only acceptable modern DAC chip is PCM1794.
Which is, coincidentally, 6-bit R2R......
An externally hosted image should be here but it was not working when we last tested it.
Do you have some datasheet reference that points to the internal DAC being R2R? For 6 bits, hardly seems worth it.
I really think this needs to be rewritten on Wikipedia.
Audio bit depth - Wikipedia, the free encyclopedia
"The perceived dynamic range of 16-bit audio can be as high as 120 dB with noise-shaped dither, taking advantage of the frequency response of the human ear."
I'm pretty sure I could make a 16-bit / 44.1 kHz music file on my computer with 121 dB perceived dynamic range?
Either that or a 16-bit / 44.1 kHz file which went via ADC with zero dither and a DAC with 64x oversampling to shift the quantization noise into the ultrasonic?
Not sure about the second part.
Audio bit depth - Wikipedia, the free encyclopedia
"The perceived dynamic range of 16-bit audio can be as high as 120 dB with noise-shaped dither, taking advantage of the frequency response of the human ear."
I'm pretty sure I could make a 16-bit / 44.1 kHz music file on my computer with 121 dB perceived dynamic range?
Either that or a 16-bit / 44.1 kHz file which went via ADC with zero dither and a DAC with 64x oversampling to shift the quantization noise into the ultrasonic?
Not sure about the second part.
Nevermind I was incorrect here
Wiki said:
Thus, 1-bit is the highest volume / SPL resolution necessary.
Does this help?
An externally hosted image should be here but it was not working when we last tested it.
I can find this
http://www.diyaudio.com/forums/digital-line-level/137623-pcm1704-newer-chips-4.html#post1740069
This
"To learn more details regarding the Advanced Segment
DAC architecture, please refer to the paper presented at the
109 th AES Convention entitled “A 117db, D-Range, Cur-rent-
mode, Multi-Bit, Audio DAC for PCM and DSD Audio
Playback” by Nakao, Terasawa, Aoyagi, Terada, and
Hamasaki of Burr-Brown Japan"
This
http://www.technobase.jp/eclib/OTHER/DATASHEET/BB/dsd1792a.pdf
"This architecture has overcome various drawbacks of conventional multi-bit processing"
It's multi-bit and non-conventional.
Perhaps it's not actually R2R. In that case I've been lied to, should stop listening as usual.
If you can, check this paper
“A 117db, D-Range, Cur-rent-
mode, Multi-Bit, Audio DAC for PCM and DSD Audio
Playback”
That should settle the issue.
The fact that DWA (data-weighted averaging) is shown tends to indicate its a 'thermometer DAC' rather than R2R. Having such an architecture lends itself to turning mis-matches in individual resistors in the 'thermoneter' into noise.
- Status
- Not open for further replies.