Doesn't that boil down to: what is the lowest amplitude resolution a human can reliably hear, period? Does it matter what the source is of the amplitude difference?
Just trying to cut to the chase.
Jan
the difference between 0dB SPL and 1dB SPL.
one can be heard and the other can't be heard by most people.
But is that the lowest resolution?
Can we hear the difference between 2dBSPL and 2.5dB SPL?
Can we hear a difference when listening to a pair of two tone signals? Say 500Hz @ 20dB SPL in both signals and 20dB or 20.2dB of 1100Hz in the two test signals.
Can we detect that difference of 0.2dB, when there is a reference signal in the test?
If noise is added to all of these tests.
Would the results be the same if the noise is 30dB below the test signals?
What if the noise was raised to 10dB below the test signals?
And raised even further to equal the test signals?
one can be heard and the other can't be heard by most people.
But is that the lowest resolution?
Can we hear the difference between 2dBSPL and 2.5dB SPL?
Can we hear a difference when listening to a pair of two tone signals? Say 500Hz @ 20dB SPL in both signals and 20dB or 20.2dB of 1100Hz in the two test signals.
Can we detect that difference of 0.2dB, when there is a reference signal in the test?
If noise is added to all of these tests.
Would the results be the same if the noise is 30dB below the test signals?
What if the noise was raised to 10dB below the test signals?
And raised even further to equal the test signals?
I agree. The observer has not been defined previously, so a limitless observer was assumed, which is incorrect. Our [avarage?] Hearing ought to be established first....and modelled.
Hey everyone,
Here is the 4-bit, zero-dither file ----- http://www.audiochrome.net/clips/Venice_4b_nodither.mp3
Here is the 4-bit, zero-dither file ----- http://www.audiochrome.net/clips/Venice_4b_nodither.mp3
Your delta sigma comparison to a R2R if faulty as both FAIL TO INCLUDE THE RECONSTRUCTION FILTER which is NOT an optional part of the process.
Apply the recostruction filter and both waveforms will turn into a clean sine wave.
Further your question is badly formed as it ignores the ability to raise the sample rate and shorten the word length without loosing any information, raise the rate 4 times and you can reduce the word length by 1 bit while preserving all the signal content (This is part of what that Delta Sigma dac is doing).
It is easier to build a short wordlength dac running at a high rate then it is to build a 24 bit dac running at 44.1, so an upsampler followed by a (correctly dithered) short DAC is often better then a full length simple minded DAC, there is a reason modern parts need ~25MHz clocks....
Seriously, you need to read a few theory books to even be able to form the question in an answerable way.
Human threshold of hearing is ~20uPa which we define as 0dBSpl, thus in a theoretical sense in a perfect room the required dynamic range would be defined by how far above 0dBSPL we wished to be able to go.
Now a real room (Even the very quiet aechoic chamber at work) seldom measures less then 20dBSPL, and my listening room clocks in at around 30dBSPL or so (Mainly road noise), and I for one don't really like listening at 120dBSPL (Pain threshold, also hearing damage), so by subtracting one from the other I get that ~90dB dynamic range is quite sufficient hence ~16 bits is sufficient in a repro chain.
Regards, Dan.
Apply the recostruction filter and both waveforms will turn into a clean sine wave.
Further your question is badly formed as it ignores the ability to raise the sample rate and shorten the word length without loosing any information, raise the rate 4 times and you can reduce the word length by 1 bit while preserving all the signal content (This is part of what that Delta Sigma dac is doing).
It is easier to build a short wordlength dac running at a high rate then it is to build a 24 bit dac running at 44.1, so an upsampler followed by a (correctly dithered) short DAC is often better then a full length simple minded DAC, there is a reason modern parts need ~25MHz clocks....
Seriously, you need to read a few theory books to even be able to form the question in an answerable way.
Human threshold of hearing is ~20uPa which we define as 0dBSpl, thus in a theoretical sense in a perfect room the required dynamic range would be defined by how far above 0dBSPL we wished to be able to go.
Now a real room (Even the very quiet aechoic chamber at work) seldom measures less then 20dBSPL, and my listening room clocks in at around 30dBSPL or so (Mainly road noise), and I for one don't really like listening at 120dBSPL (Pain threshold, also hearing damage), so by subtracting one from the other I get that ~90dB dynamic range is quite sufficient hence ~16 bits is sufficient in a repro chain.
Regards, Dan.
Yes, we are now entering bizzaro-land.
Where people wear shoes as hats, hats as shoes, tie laces on their hands.
Where snow falls up, heat falls down, night rises in the day, day sets in the night.
Where resolution is exclusively within time.
Time is exclusively within resolution.
Where 4-bit sounds high-end, if you remove the quantization noise.
Is that so?
If we used some sophisticated, advanced algorithm to remove the quantization noise?
Or if we just removed it pixel by pixel?
Does it sound high-end then?
Does it?
______
For real, if we exclude dynamic range, is everything we need in 4-bit / 44.1 kHz audio, if we can or could efficiently remove the quantization noise?
Yes I'm naïve, if this is true I didn't know that.
I have tested the file with my Nos DAC and ES9018 DAC, normal play and 2x upsampling.
I have deactivated 64-bit ASIO in Foobar, I have used a different media player in 8-bit Integer precision.
I can not hear any noteworthy difference anywhere, this file sounds pretty much the same everywhere, with or without the 2x upsampling.
Last edited:
Take that 4 bit converter and do the following:
First upsample your signal by 256 times then dither to 4 bits and truncate, that 4 bit DAC now has the noise floor of an 8 bit dac within the audio band, the out of band stuff will be removed by the reconstruction filter (That can now go over at 50K or so as the upsample process has put the first image 256 times further away then in the NOS case.
Now noise shape the dither so that almost all of the energy appears above the reconstruction filters cutoff, that might get us another 24dB or so, so maybe another 4 bits on the effective noise floor, 12 bits. Not quite high end in my view, but perfectly acceptable for some things (The BBC used 13 bits for their studio - transmitter links for many years and no radio 3 listener ever complained).
I figure that starting with a 6 bit DAC running at maybe 50MHz you could get 16 bits in the audio band without too much pain and without doing the delta-sigma thing, just using upsampling and noise shaped dither.
Regards, Dan.
First upsample your signal by 256 times then dither to 4 bits and truncate, that 4 bit DAC now has the noise floor of an 8 bit dac within the audio band, the out of band stuff will be removed by the reconstruction filter (That can now go over at 50K or so as the upsample process has put the first image 256 times further away then in the NOS case.
Now noise shape the dither so that almost all of the energy appears above the reconstruction filters cutoff, that might get us another 24dB or so, so maybe another 4 bits on the effective noise floor, 12 bits. Not quite high end in my view, but perfectly acceptable for some things (The BBC used 13 bits for their studio - transmitter links for many years and no radio 3 listener ever complained).
I figure that starting with a 6 bit DAC running at maybe 50MHz you could get 16 bits in the audio band without too much pain and without doing the delta-sigma thing, just using upsampling and noise shaped dither.
Regards, Dan.
O.k., thank you for your answers by the way D. Mills.
Your above post I need to think about the technical part.
But this answers my question, doesn't it?
I played an undithered 4-bit / 44.1 kHz file in a NOS DAC and the only thing missing is dynamic range and the only extra present is quantization noise.
Is this true? Really?
Yes you are right by the way, resolution is a confusing term within audio. What does a soft tap and a loud tap have to do with resolution? Nothing, really, in the way the vast majority of people imagine the concept of resolution, like a low resolution video versus a high resolution video.
I thought the broken-ness of a 4-bit sine would cause THD? Versus a pure sine?
Your above post I need to think about the technical part.
But this answers my question, doesn't it?
I played an undithered 4-bit / 44.1 kHz file in a NOS DAC and the only thing missing is dynamic range and the only extra present is quantization noise.
Is this true? Really?
Yes you are right by the way, resolution is a confusing term within audio. What does a soft tap and a loud tap have to do with resolution? Nothing, really, in the way the vast majority of people imagine the concept of resolution, like a low resolution video versus a high resolution video.
I thought the broken-ness of a 4-bit sine would cause THD? Versus a pure sine?
Last edited:
And no. 4 bits at 44.1 WILL (If it was done correctly) have a broadband noise floor at the -21dB level, that is what 4 bits means!
There is no way to remove the quantisation noise, the best that can be done is to move it around the spectrum to minimise its audiability.
8 Bit integer should be noticable as additional hiss in the background (But if the conversion was done correctly, that will be all).
64 bits is nonsense for an audio DAC, physically meaningless.
I would have been very surprised if upsampling by a factor of two made any difference to what you heard, it adds no new information.
Go read a signal processing book, then the (generally very good) white papers from people like Dan Larvy, Texas Instruments, Crystal Semi....
We cannot teach you the basics online, that needs a book.
Regards, Dan.
There is no way to remove the quantisation noise, the best that can be done is to move it around the spectrum to minimise its audiability.
8 Bit integer should be noticable as additional hiss in the background (But if the conversion was done correctly, that will be all).
64 bits is nonsense for an audio DAC, physically meaningless.
I would have been very surprised if upsampling by a factor of two made any difference to what you heard, it adds no new information.
Go read a signal processing book, then the (generally very good) white papers from people like Dan Larvy, Texas Instruments, Crystal Semi....
We cannot teach you the basics online, that needs a book.
Regards, Dan.
I can recommend: 'Introductory Digital Signal Processing' by Lynn and Fuerst. Reading it as we speak. Sort of.
Jan
Jan
As a theoretical concept, if we could remove the quantization noise and if we did not care about dynamic range very much, then 4-bit / 44.1 kHz audio would be perfectly sufficient, for normal consumer use.
This is what I'm arriving at.
No, 4-bits means dynamic range, that's all it means!
Applying shaped dither is simply a human evaluation of what performs / sounds better.
Quantization noise has no self-existing noise floor, when the music passage becomes truly silent, there is true silence.
Nope for a quantiser to be linear it MUST BE DITHERED, we have a word for an undithered quantiser, that word is "broken".
If you have dither then you have a defined broadband noise floor, which (Best case) sets the broadband dynamic range, if the noise floor is at -90dBFS then the dynamic range is clearly 90dB.
The trick is that you can use a bandwidth wider then the audio band and contrive to place most of the energy required for the dither where it is less audiable, so you can have a dynamic range in a bandwidth narrower then Fs/2 that is greater then the full bandwidth dynamic range.
Further if the converter is running at a rate a large multiple of the required bandwidth then the reconstruction filter can effectively average several samples allowing a futher improvement (This is where the extra bit with every 4 times increase in SR comes from).
Trust me, go read a book on the theory, this stuff is not intuitive.
Regards, Dan.
If you have dither then you have a defined broadband noise floor, which (Best case) sets the broadband dynamic range, if the noise floor is at -90dBFS then the dynamic range is clearly 90dB.
The trick is that you can use a bandwidth wider then the audio band and contrive to place most of the energy required for the dither where it is less audiable, so you can have a dynamic range in a bandwidth narrower then Fs/2 that is greater then the full bandwidth dynamic range.
Further if the converter is running at a rate a large multiple of the required bandwidth then the reconstruction filter can effectively average several samples allowing a futher improvement (This is where the extra bit with every 4 times increase in SR comes from).
Trust me, go read a book on the theory, this stuff is not intuitive.
Regards, Dan.
Are the Oscilloscope traces are a good representation of reality?
How does the output from a known good source like an analog oscillator look on the Oscilloscope with the same settings?
What does it look like on an old analog oscilloscope.
How does the output from a known good source like an analog oscillator look on the Oscilloscope with the same settings?
What does it look like on an old analog oscilloscope.
I was looking at that 4-bit music file with the oscilloscope in Foobar just now and everything looks perfectly normal.
You can't see very much in oscilloscope waves, but that's just my IMHO there.
OK I agree the Oscilloscope waveforms aren't the full story. I was referring to the trace showing the HM-602 Headphones-out which doesn't look right. Either the wobble has been created by the Oscilloscope (it may be only 8 bit) or the DAC output is poor or the amplitude is very low.
He calls it step distortion, I just came up with stair-case-distortion now.
You and he are of the same creed, you both think Non-Interpolating DAC's sound more natural.
Altmann says it's due to linarity, rather than all the ultrasonic noise which is present within Delta-Sigma.
He says a 4x AD1955 is the best Delta-Sigma he's heard.
You say it's due to glitch error and perhaps due to distortion spectrum character as well, if I've read you correctly.
XXHighEnd says it's due to perfection within the time domain and his corrective method to achieve 0.001% THD.
Last edited:
"For a quantizer to be linear"
Technically speaking, only reality is linear. A quantizer, with or without dither, is non-linear.
There is no ADC which has perfectly sampled reality, if you think there is just let me know.
If you want to skip the noise floor part, don't use dither, then we arrive at the true meaning of audio bit-depth, which is dynamic range.
The mathematical implication to use dither, is to re-approximate reality.
If there's some kind of evidence that an ADC is truly linear, with reality as the reference point, just show me the link and I'll be quiet.
Technically speaking, only reality is linear. A quantizer, with or without dither, is non-linear.
There is no ADC which has perfectly sampled reality, if you think there is just let me know.
If you want to skip the noise floor part, don't use dither, then we arrive at the true meaning of audio bit-depth, which is dynamic range.
The mathematical implication to use dither, is to re-approximate reality.
If there's some kind of evidence that an ADC is truly linear, with reality as the reference point, just show me the link and I'll be quiet.
- Status
- Not open for further replies.