a/d converter chips

[ricardo]

http://src.infinitewave.ca/

Thanks for this abbey.  It's the main article I was trying to find.  Some new SRCs have been added but the only new member of the elite group is Apple CoreAudio(Leopard)  Have a look at this to see what good SRC should be like.  No visible aliasing and the intermod this introduces down to -150dB.  Only Logic 8 (Leopard) seems to have taken advantage of this.

So the designers make their choices. Some privilege transient response at the expense of out-of-band attenuation or HF extension, some privilege roll-off at the risk of being criticized for the harshness of cymbals ...

Actually hardness of cymbals is caused by not having enough roll-off cos the aliasing and intermod.

I've conducted Double Blind Listening Tests bla bla on several of the tested parameters for other purposes and will put money that some of my panel can reliably detect the aliasing and intermod on certain (musical) signals.  These people (the best ears I know) will not be able to detect some of the other stuff like transient response except under highly artificial conditions.

On the subject of real-time SRC, don't we now have enough processing power to upsample to the Lowest Common Multiple?  Excuse me if this is a stupid question but my (pseudo) DSP knowledge is firmly 20th century.

Rochey, could you give us a firm (or official) opinion?  Which is TI's best A/D?

My best DA converter so far gets to 130 dB dynamic range in a 22 kHz measurement bandwidth.

Sam, is this your measurement or a claim?  What device is this?

One nuance or subtlety surrounding dynamic range discussions for A/D/A is the classic analog definition (peak signal to noise floor) ignores what is going on below this nominal noise floor. We can hear signals that exist below this nominal noise floor so there is another signal floor down at the quantization limit.  How far this quantization occurs below the noise is probably audible if close.

I can answer that.  The Sony PCM-F1 had textbook 16b dithered performance.  You can take a sine wave 20dB below the noise and still hear (after suitable amplification) a clean if noisy sine wave.  If you switch the F1 to 14b EIAJ, it is improperly dithered and you get the crackling, loss of signal and noise modulation well before you reach the noise level.  Piano music is good for this test too.

I used this in da early days to prove to others (& myself) that this digital stuff was pukka.

This is a simple test for a 16b converter not needing any gear except an attenuator, hi-gain line amplifier and your ears.  IMHO, a codec which didn't do this is unacceptable.  Unfortunately, if you look at the Rightmark website, you see many soundcards & recorders that would fail.

Some commonly available examples of good 16 codecs today are the Line I/Ps on IBM / Lenovo Thinkpads going back to at least the T22.

In theory, low distortion quantisation (visible by averaging over a zillion seconds) is available a zillion dB below the properly dithered noise level but piano -20dB below noise is a good practical benchmark.

 

[JohnRoberts]

One nuance or subtlety surrounding dynamic range discussions for A/D/A is the classic analog definition (peak signal to noise floor) ignores what is going on below this nominal noise floor. We can hear signals that exist below this nominal noise floor so there is another signal floor down at the quantization limit.  How far this quantization occurs below the noise is probably audible if close.

I can answer that.  The Sony PCM-F1 had textbook 16b dithered performance.  You can take a sine wave 20dB below the noise and still hear (after suitable amplification) a clean if noisy sine wave.  If you switch the F1 to 14b EIAJ, it is improperly dithered and you get the crackling, loss of signal and noise modulation well before you reach the noise level.  Piano music is good for this test too.

I used this in da early days to prove to others (& myself) that this digital stuff was pukka.

This is a simple test for a 16b converter not needing any gear except an attenuator, hi-gain line amplifier and your ears.  IMHO, a codec which didn't do this is unacceptable.  Unfortunately, if you look at the Rightmark website, you see many soundcards & recorders that would fail.

Some commonly available examples of good 16 codecs today are the Line I/Ps on IBM / Lenovo Thinkpads going back to at least the T22.

In theory, low distortion quantisation (visible by averaging over a zillion seconds) is available a zillion dB below the properly dithered noise level but piano -20dB below noise is a good practical benchmark.

To maintain 20 dB of signal integrity below the nominal noise floor requires a little more than 3 bits, so a 24b convertor with 20 bit S/N seems in the ballpark, and should be adequately "analog" sounding down in the dirt.

JR

PS: Still no thoughts on complaints about digital summing buses? I have been looking for a smoking gun there for years. Just anecdotal poop. 

 

[ricardo]

To maintain 20 dB of signal integrity below the nominal noise floor requires a little more than 3 bits, so a 24b convertor with 20 bit S/N seems in the ballpark, and should be adequately "analog" sounding down in the dirt.

This is a common misconception even among pseudo DSP gurus.  My experiments were conducted with properly dithered 16b codecs.  ie these systems have practical resolution well below their noise floor and are indistinguishable from a perfect band-limited analogue chain with the same S/N.

The proper random dither provides "infinite resolution sampling".  You don't need extra bits to do this.

Record a low distortion sine wave at -113dB fs (20dB below noise) on the (properly dithered 16b) F1 and average the result (to reduce the noise) until you recovered the sine wave.  You would find it is still low distortion.

BTW, the noise of a properly TPDF dithered 16b channel over 20kHz is -93.32dB wrt to a fs sine wave.  It is NOT 16x6.03dB.  The corresponding figures for 20b & 24b are -117.4dB & -141.49dB.  See Lipsh*tz & Vanderkooy JAES articles for chapter & verse.  If you measure a 16b chain with better than -93dB 20kHz noise wrt to a fs sine, it is improperly dithered and will exhibit digital distortion.

PS: Still no thoughts on complaints about digital summing buses? I have been looking for a smoking gun there for years. Just anecdotal poop.

Du..uh! dunno.  I'm only a pseudo DSP guru

 

[JohnRoberts]

To maintain 20 dB of signal integrity below the nominal noise floor requires a little more than 3 bits, so a 24b convertor with 20 bit S/N seems in the ballpark, and should be adequately "analog" sounding down in the dirt.

This is a common misconception even among pseudo DSP gurus.  My experiments were conducted with properly dithered 16b codecs.  ie these systems have practical resolution well below their noise floor and are indistinguishable from a perfect band-limited analogue chain with the same S/N.

The proper random dither provides "infinite resolution sampling".  You don't need extra bits to do this.

I delayed responding to this hoping a real digital expert would chime in.

My understanding of how dither works is somewhat less optimistic. The (amplitude) resolution that dither can can add seems limited by the fact that it is quantized along the time axis by sample length. So that LSB can be dithered up/down by noise (or shaped noise, or whatever dither), but it remains stable up or down for one full sample at a time. Creating resolution below this LSB is effected by alternate patterns of up/down, just like PWM. 1up/1down=50% or one more bit down, 1 up/3 down is 2 bits down, 1 up/7 down is 3 bits down etc. We clearly run into sample rate/vs signal frequency issues long before gaining more than a few bits of extra resolution, while those few extra bits are surely worthwhile especially for low bit rate systems, compared to a nasty noise floor with quantization distortion.

My understanding that modern high performance convertors have actual analog (like) noise floors several bits higher than their quantization floor stands, while I don't claim that the noise source is simple, but related to the oversampling and decimation (perhaps from the initial fast/low bit conversion). This few bits higher level of noise, makes dithering the LSB with LSB level noise academic. (In my judgement).

But don't take my word for it... Hopefully an expert will chime in and tell us which analog guy is less wrong.. 

JR

 

[Samuel Groner]

Sam, is this your measurement or a claim?

Measurement, of course!

Samuel

 

[ricardo]

My understanding that modern high performance convertors have actual analog (like) noise floors several bits higher than their quantization floor stands, while I don't claim that the noise source is simple, but related to the oversampling and decimation (perhaps from the initial fast/low bit conversion). This few bits higher level of noise, makes dithering the LSB with LSB level noise academic.

John this is indeed the case with present 24b converters.  The only 24b converter where dither is at the LSB level is the experimental device Scott Wurcer tested.  All others (good ones) are noise limited at about the 20b level though from what Sam says, this is slowly improving.

Samuel, which of your A/Ds did you measure at 130dB dynamic range?

But there are a few good 16b converters where dither is 'at' (see L&V for exact details) the LSB level.  The Sony PCM-F1 is done straight w/o oversampling etc.  The few IBM Thinkpads I've measured are also good.  Probably oversampling.  There are a lot more 16b converters which don't come anywhere near theoretical 16b performance.  If you were cynical, you could claim ALL 24b converters are properly dithered!

The issues of linearity below the LSB dither noise level are discussed by Lipshitz & Vanderkooy in at least 3 JAES articles.  As with anything of the Dynamic Duo, no further elaboration is possible.

But don't take the DD's or my word for it.  Test it for yourself.  Record something at -113dB wrt FS on a good 16b system.  You don't need to use averaging, sine waves and a distortion meter.  Music will do.  Then compare with a noisy analog chain.  (You can make this by simply having a large series resistor into a high gain high Z amplifier.)  All this needs is an attenuator, a high gain line level amplifier and your ears.

 

[abbey road d enfer]

On the subject of real-time SRC, don't we now have enough processing power to upsample to the Lowest Common Multiple?

Unfortunately, in real-time, the ratio is essentially irrational, so the algo has to take the closest rational ratio, which may lead to a very high LCM.
Just take the example of converting from 48 khz to 44.101, the LCM is 211.6848MHz, which would leave only 1.5 instruction cycle on the fastes DSP alive. But dealing with 44.1001 kHz shows it just is not feasible, going into the GHz range.
Nor is it really necessary; the algos have a limited upsampling range and whatever overflow is generated there is taken care of by buffers.