Please also consider that p-p jitter is generally a statistical measurement, and without a specified BER is very difficult to use.
Hi,
Thanks for the reply, most interesting.
Here is where I found those calculations
http://www.lcaudio.com/intlcl2.htm
They use 2^20 for the divisor and I agree with you that it should be 2^16 for CD even when oversampled.
Maybe they are just trying to promote thier products. Although I have heard the same theory else where.
Here is another link which provides some interesting info on digital filters.
http://www.iar-80.com/
Go to the bottom of the page. Start with the DCS Purcell review and then goto the Background Analysis Page.
The Author makes some interesting points about oversampling filters (Impluse Response) and Noise Shaping.
Thanks for the reply, most interesting.
Here is where I found those calculations
http://www.lcaudio.com/intlcl2.htm
They use 2^20 for the divisor and I agree with you that it should be 2^16 for CD even when oversampled.
Maybe they are just trying to promote thier products. Although I have heard the same theory else where.
Here is another link which provides some interesting info on digital filters.
http://www.iar-80.com/
Go to the bottom of the page. Start with the DCS Purcell review and then goto the Background Analysis Page.
The Author makes some interesting points about oversampling filters (Impluse Response) and Noise Shaping.
background article
I read through most of the article. My impression is that the author knows a great many details, but he is thoroughly misguided on a couple of concepts. I have not seen his example on a V vs. W reconstruction, but I believe it could be an artefact from filter ringing.
Some points where he is clearly in error:
- aliasing will add additional energy and hence sounds in the <20 kHz range
- ultrasonic energy, when unfiltered, may cause slewing distortions in the preamp and amp
- according to the sampling theorem, two samples per period (i.e. full sine) are enough to reproduce a signal, low-pass filtering to 1/2 the sampling rate will reproduce the original signal, no matter whether done in the digital or analog domain
- the averaging he describes does not involve drawing a straight line between two samples
- a digital filter does not copy the samples 4 or 8 times, rather it will include (n-1) zero samples, where n is the oversampling ratio
- low pass filtering will smooth the curve, this is in effect some sort of averaging or interpolation
- he is right in that those interpolated or smoothed samples will have a higher resolution, truncating them to 16 bits will introduce some additional noise
- he is wrong in saying that after all those oversamling operations, the sample could be cut back to 16 bit/44.1 kHz. in that case, we would undo the oversampling action and be back in situation where we have two samples per period that need low-pass filtering to reconstruct the sine
I stopped reading somewhere on the continued page because all those well-meant misconceptions were beginning to unnerve me.
Eric
I read through most of the article. My impression is that the author knows a great many details, but he is thoroughly misguided on a couple of concepts. I have not seen his example on a V vs. W reconstruction, but I believe it could be an artefact from filter ringing.
Some points where he is clearly in error:
- aliasing will add additional energy and hence sounds in the <20 kHz range
- ultrasonic energy, when unfiltered, may cause slewing distortions in the preamp and amp
- according to the sampling theorem, two samples per period (i.e. full sine) are enough to reproduce a signal, low-pass filtering to 1/2 the sampling rate will reproduce the original signal, no matter whether done in the digital or analog domain
- the averaging he describes does not involve drawing a straight line between two samples
- a digital filter does not copy the samples 4 or 8 times, rather it will include (n-1) zero samples, where n is the oversampling ratio
- low pass filtering will smooth the curve, this is in effect some sort of averaging or interpolation
- he is right in that those interpolated or smoothed samples will have a higher resolution, truncating them to 16 bits will introduce some additional noise
- he is wrong in saying that after all those oversamling operations, the sample could be cut back to 16 bit/44.1 kHz. in that case, we would undo the oversampling action and be back in situation where we have two samples per period that need low-pass filtering to reconstruct the sine
I stopped reading somewhere on the continued page because all those well-meant misconceptions were beginning to unnerve me.
Eric
I'm a little late at this ... but: you should look at this:
http://sound.westhost.com/project85.htm
I'm starting to build one now.....
It will fit your need perfectly I think....
http://sound.westhost.com/project85.htm
I'm starting to build one now.....
It will fit your need perfectly I think....
capslock:
you raise some good points re. jitter sensetivity of oversampled streams and I-V conversion slewing errors. It looks like i was actually thinking about both slew rate limitations and settling time simultaneously and just hadn't taken the time to think it all through carefully and properly compose my last response. In any case, you are probably correct that the aforementioned slew rate limitation is likely negated by single large steps being replaced with several smaller steps. However, overshoot and settling times may not be linearly affected by the size of the step, particularly if the step is fast enough to put the opamp into a slewing condition. So, one might still expect to see some variation in I-V converter performance... whereby with a greater number of slewing steps, the opamp spends a greater proportion of time in overshoot and ringing rather than settled at the correct voltage. Then again, if the ringing is symmetrical (no DC component of it's own), then this should not matter, as the overshoot and ringing frequency components will be averaged out to the correct output value by subsequent low-pass filtering.
Anyway, all previous discussion aside, I have *finally* found a technical paper which explains to my satisfaction why some digital audio oversampling filters may be degrading the audio quality. The paper is here:
http://www.nanophon.com/audio/antialia.pdf
I found this paper particularly insightful, and I'm sure you will too. Julian Dunn also has a number of other excellent articles on his site. ( http://www.nanophon.com/audio/index.htm ) Have a gander, and let me know what you think!
you raise some good points re. jitter sensetivity of oversampled streams and I-V conversion slewing errors. It looks like i was actually thinking about both slew rate limitations and settling time simultaneously and just hadn't taken the time to think it all through carefully and properly compose my last response. In any case, you are probably correct that the aforementioned slew rate limitation is likely negated by single large steps being replaced with several smaller steps. However, overshoot and settling times may not be linearly affected by the size of the step, particularly if the step is fast enough to put the opamp into a slewing condition. So, one might still expect to see some variation in I-V converter performance... whereby with a greater number of slewing steps, the opamp spends a greater proportion of time in overshoot and ringing rather than settled at the correct voltage. Then again, if the ringing is symmetrical (no DC component of it's own), then this should not matter, as the overshoot and ringing frequency components will be averaged out to the correct output value by subsequent low-pass filtering.
Anyway, all previous discussion aside, I have *finally* found a technical paper which explains to my satisfaction why some digital audio oversampling filters may be degrading the audio quality. The paper is here:
http://www.nanophon.com/audio/antialia.pdf
I found this paper particularly insightful, and I'm sure you will too. Julian Dunn also has a number of other excellent articles on his site. ( http://www.nanophon.com/audio/index.htm ) Have a gander, and let me know what you think!
Slew rate limitations
The way around that is to not use feedback.
I am not sure a MOSFET is the best choice, but a true transimpedance amp is not that hard to make.
And don't believe everything Wadia is telling you.
Remember, BJTs are current controlled devices.
Glad to see CAPSLOCK still agrees with me.
Jocko
The way around that is to not use feedback.
I am not sure a MOSFET is the best choice, but a true transimpedance amp is not that hard to make.
And don't believe everything Wadia is telling you.
Remember, BJTs are current controlled devices.
Glad to see CAPSLOCK still agrees with me.
Jocko
HifiZen:
This is a very good article, only drawback being that it explores only equi-ripple designs (DF1704, PCM1738, AD185x are not equi-ripple) but I guess to some degree the findings can be generalized.
The Japanese guy who started the no-oversampling crusade said that long filters were bad because the timing information was smeared over a greater intervall, and the longer the intervall, the more likely it was to be audible.
Julian Dunn accedes that the separation of the original impulse and its artefacts becomes greater if the filter has more taps (i.e. is longer). This is obvious to anybody who has designed FIR filters. However, he proves that in an equiripple filter, there is no smearing, rather there are two distinct echoes. The flatter the ripple, the weaker the echoes become. This is again obvious, because a good filter will have a long, but gentle inpulse response, i.e. it will have many coefficients but those start with very small values. The main point is: pre and post echoes at +/- 1 ms and -100 dB are probably preferable to +/- 0.3 ms at -25 dB.
On a side note, personally, I wonder why there are so few IIR filters around...
Jocko:
I am not sure I agree with your last post. A stage with degeneration (i.e. local current feedback) may well be less susceptible to slewing distortions than a classic global voltage feedback op amp.
But: With modern op amps you should not come anywhere near the slewing limit. As Julian Dunn points out (implicitly), for alias images to mirror into the pass band, all you need is a static nonlinearity of the transfer function. A single stage may actually be more non-linear than a decent op amp.
Eric
This is a very good article, only drawback being that it explores only equi-ripple designs (DF1704, PCM1738, AD185x are not equi-ripple) but I guess to some degree the findings can be generalized.
The Japanese guy who started the no-oversampling crusade said that long filters were bad because the timing information was smeared over a greater intervall, and the longer the intervall, the more likely it was to be audible.
Julian Dunn accedes that the separation of the original impulse and its artefacts becomes greater if the filter has more taps (i.e. is longer). This is obvious to anybody who has designed FIR filters. However, he proves that in an equiripple filter, there is no smearing, rather there are two distinct echoes. The flatter the ripple, the weaker the echoes become. This is again obvious, because a good filter will have a long, but gentle inpulse response, i.e. it will have many coefficients but those start with very small values. The main point is: pre and post echoes at +/- 1 ms and -100 dB are probably preferable to +/- 0.3 ms at -25 dB.
On a side note, personally, I wonder why there are so few IIR filters around...
Jocko:
I am not sure I agree with your last post. A stage with degeneration (i.e. local current feedback) may well be less susceptible to slewing distortions than a classic global voltage feedback op amp.
But: With modern op amps you should not come anywhere near the slewing limit. As Julian Dunn points out (implicitly), for alias images to mirror into the pass band, all you need is a static nonlinearity of the transfer function. A single stage may actually be more non-linear than a decent op amp.
Eric
Can't remember why IIR filters aren't used........I'm having a senior-citizen moment. But whatever the reason, it was a good one.
Yes, a good modern op-amp should not have slew problems, but which manufacturer uses them in production.
I've tried voltage-feedback, current-feedback, and a simple transimpedance amp with no feedback.
I much prefer the latter. (Probably because I am not a big fan of global feedback biases my belief.)
Artifacts make for IM products, and ICs are more susceptible to out-of-band stuff than discrete. Just as they are more susceptible to EMI-induced IMD.
And non-o/s has lots of artifacts.
Jocko
Yes, a good modern op-amp should not have slew problems, but which manufacturer uses them in production.
I've tried voltage-feedback, current-feedback, and a simple transimpedance amp with no feedback.
I much prefer the latter. (Probably because I am not a big fan of global feedback biases my belief.)
Artifacts make for IM products, and ICs are more susceptible to out-of-band stuff than discrete. Just as they are more susceptible to EMI-induced IMD.
And non-o/s has lots of artifacts.
Jocko
<i>"Can't remember why IIR filters aren't used"</i>
most of the DFs are built to run on 32k 44k1 48k and maybe higher, is it harder to impliment multirate IIR ?
From memory the digital de-emphisis on some chips use IIR
<hr width="95%" align=center>
Most of the ZeroOS DACs are trying to ape the Fluency conversion process that Luxman released some years ago, a kind of Curve fitting, unfortunatly I haven't read a good explaination of process.
Regards
James
most of the DFs are built to run on 32k 44k1 48k and maybe higher, is it harder to impliment multirate IIR ?
From memory the digital de-emphisis on some chips use IIR
<hr width="95%" align=center>
Most of the ZeroOS DACs are trying to ape the Fluency conversion process that Luxman released some years ago, a kind of Curve fitting, unfortunatly I haven't read a good explaination of process.
Regards
James
I seem to recall it has something to do with low-level instabilites. That's all I can rememeber.
Jocko
Jocko
- Status
- Not open for further replies.