-
Welcome to the General Discussion forum for UAD users! Please note that this forum is user-run, although we're thrilled to have so much contribution from Drew, Will, and other UA folks! Feel free to discuss both UAD and non-UAD related subjects! 1) Please do not post technical issues here. Please use our UAD Support Forums instead. 2) Please do not post complaints here. Use the Unrest Forum instead. They have no place in the the General Discussion forum. Threads posted in the wrong forum will be moved, so if you don't see your thread here anymore, please look in the correct forum. Lastly, please be respectful.
P Limiter needs update to catch intersample overs!
- Thread starter seti808
- Start date
rimbaud234
Member
Intersample overs are a tricky beast. There was a great panel on it at AES in NYC. It's actually not an easy problem to define or solve.
It depends on the characteristics of your D/A for one thing.
Also, would you be willing to take the CPU hit?
And the big question is when have you had big problems with intersample overs?
It depends on the characteristics of your D/A for one thing.
Also, would you be willing to take the CPU hit?
And the big question is when have you had big problems with intersample overs?
Dan Duskin
Established Member
i think this was discussed in an older webzine... wasn't it?
http://www.uaudio.com/webzine/2005/sept ... ndex2.html
...As an aside, a common topic related to resampling is peak detection in limiting, which is based on the \"interpolated peaks\" of a signal. Under this scheme, resampling is used in an attempt to discover the peak levels of the analog signal coming out of the D/A converters, rather than using the peak values of the signal samples themselves. This is thought to be more conservative, given that the analog headroom of the D/A is unknown. It is interesting to note that, using perfect signal reconstruction, it is possible to construct a signal whose samples all lie between plus and minus 1.0, but whose interpolated peak grows in an unbounded way as the length (in time) of the signal grows. This means that it is technically impossible to achieve true limiting on interpolated peaks, unless the peak limiter is allowed to have zero gain, in which case there is no output!...
...As an aside, a common topic related to resampling is peak detection in limiting, which is based on the \"interpolated peaks\" of a signal. Under this scheme, resampling is used in an attempt to discover the peak levels of the analog signal coming out of the D/A converters, rather than using the peak values of the signal samples themselves. This is thought to be more conservative, given that the analog headroom of the D/A is unknown. It is interesting to note that, using perfect signal reconstruction, it is possible to construct a signal whose samples all lie between plus and minus 1.0, but whose interpolated peak grows in an unbounded way as the length (in time) of the signal grows. This means that it is technically impossible to achieve true limiting on interpolated peaks, unless the peak limiter is allowed to have zero gain, in which case there is no output!...
There is only one common method to estimate interpolated peaks in DSP.
What is done in this method is to compute values for the \"perfectly
bandlimited signal\" which passes through the sample points. This
entails designing an approximation to a perfect brick-wall filter.
The ideal brick-wall filter takes infinite computational power to
implement, which is why an approximation must be used. The
problem with the approximation is that, no matter how much
DSP is used to do the approximation, there is no way to
guarantee that the error between the peak estimate and the
actual bandlimited peak value is finite. In other words, the
error in the peak estimate can always be unbounded, no matter
how good the approximation is on the brick-wall filter.
Another problem is that there are no D/A converters that have
perfect \"reconstruction filters\". Therefore, the peak level of the
analog signal produced by the D/A will not be the same as
the peak level for the perfectly bandlimited signal. Different
D/A converters will have different filters, which will produce
different analog peak levels, for the same digital signal.
Finally, the biggest differences between interpolated peak levels
and peak sample values occur during transients. Human beings
are not very perceptive of distortion that lasts less than 4-5 mS,
so that an occasional transient which is mangled by a poor D/A
may not even be perceived. This explains the popularity of limiters
with very fast attack and release settings to take one or two dB
off transients. The reason these limiters are \"transparent\" is not
because they are not limiting very much ; it is because the time
over which the gain is reduced is very short, since only transients
reach the top one or two dB of the original signal.
To summarize: Interpolated peak estimation as currently known in the
DSP community is an approximation, with unbounded error, which
estimates something that may not resemble what comes out of
our D/A converters.
What is done in this method is to compute values for the \"perfectly
bandlimited signal\" which passes through the sample points. This
entails designing an approximation to a perfect brick-wall filter.
The ideal brick-wall filter takes infinite computational power to
implement, which is why an approximation must be used. The
problem with the approximation is that, no matter how much
DSP is used to do the approximation, there is no way to
guarantee that the error between the peak estimate and the
actual bandlimited peak value is finite. In other words, the
error in the peak estimate can always be unbounded, no matter
how good the approximation is on the brick-wall filter.
Another problem is that there are no D/A converters that have
perfect \"reconstruction filters\". Therefore, the peak level of the
analog signal produced by the D/A will not be the same as
the peak level for the perfectly bandlimited signal. Different
D/A converters will have different filters, which will produce
different analog peak levels, for the same digital signal.
Finally, the biggest differences between interpolated peak levels
and peak sample values occur during transients. Human beings
are not very perceptive of distortion that lasts less than 4-5 mS,
so that an occasional transient which is mangled by a poor D/A
may not even be perceived. This explains the popularity of limiters
with very fast attack and release settings to take one or two dB
off transients. The reason these limiters are \"transparent\" is not
because they are not limiting very much ; it is because the time
over which the gain is reduced is very short, since only transients
reach the top one or two dB of the original signal.
To summarize: Interpolated peak estimation as currently known in the
DSP community is an approximation, with unbounded error, which
estimates something that may not resemble what comes out of
our D/A converters.
That would pretty much bring this discussion to a screaching halt! Anyone willing to follow up on this one.
I have been asking myself the same thing that James is mentioning. People have told me that we don't hear distorsion that is shorter than about 1ms. So what is the deal with intersample overs, having a duration of one samle interval (about a 1/50 ms). To me it does not seem likely that you will get a whole bunch of intersample overs that will then add up to 1ms.
What I heard some limiters do, is to upsample and then limit, bringing down the sample interval to 1/100 or 1/200, but I don't see what's the big deal, you will still get intersample overs. And from what I understand, only infinate computing power would solve this.
Until someone can convince me that we can actually hear intersample distortion, I will file it under \"marketing hype\" (along with speaker cables with arrows on them). Though I must confess, it is a good one!
Petter
I have been asking myself the same thing that James is mentioning. People have told me that we don't hear distorsion that is shorter than about 1ms. So what is the deal with intersample overs, having a duration of one samle interval (about a 1/50 ms). To me it does not seem likely that you will get a whole bunch of intersample overs that will then add up to 1ms.
What I heard some limiters do, is to upsample and then limit, bringing down the sample interval to 1/100 or 1/200, but I don't see what's the big deal, you will still get intersample overs. And from what I understand, only infinate computing power would solve this.
Until someone can convince me that we can actually hear intersample distortion, I will file it under \"marketing hype\" (along with speaker cables with arrows on them). Though I must confess, it is a good one!
Petter
max cooper
Member
Buy an outboard analog limiter. It'll get 'em. If it doesn't, how will you know?
Paul Woodlock
Established Member
Cabbage said:
It's just something else to get anal about
hehe
People can't really her such short distortions.
I don't even use a limiter to raise level of my mixes, I simply increase gain, and clip the tops off those short peaks.
It sounds one helluva lot better than using a limiter/maximiser like Waves L2 etc. And I came to that conclusion by comparing at teh SAME VOLUME.
Try it yourselves if you think I'm a hethan
Paul
That's also the way I do: Cambridge with Lo-Cut Butterworth6 around 30 Hz and some narrow corrections -> LA2A or Fairchild (depending on source material) -> Precision Limiter with no more thahn 1-3db gain.Paul Woodlock said:I don't even use a limiter to raise level of my mixes, I simply increase gain, and clip the tops off those short peaks.
It sounds one helluva lot better than using a limiter/maximiser like Waves L2 etc. And I came to that conclusion by comparing at teh SAME VOLUME.
Try it yourselves if you think I'm a hethan
Paul
Loudness race is done within the compressor not the limiter. And it is much better for keeping the transients.
secretworld
Active Member
Hi everybody,
Intersample overs are a natural phenomena in d/a converters and it is a designflaw in the analog stage of cd players if they can´t handle them. But the truth is that the design flaw is there on a lot of players. There is a website wich I can´t find anymore wich has some audioexamples af a commercial (released) mix, wich you can listen to at -3dbf to keep your own converters out of the equation, with 1 the origanal mix (digtal clone) , 2 rerecorded through a converter that handels overs well and 3 a cd player (not a cheap one) that doesn´t handle overs well. The difference was very obvious!! there also was a delta file with just the distortion and that was realy shocking.
So remember it is not the distortion of the digital waveform but the clipping of the analog output stage that is the problem. But that can be serious clipping like 0.0001 at 0dbf and 10% at +1dbf
Intersample overs are a natural phenomena in d/a converters and it is a designflaw in the analog stage of cd players if they can´t handle them. But the truth is that the design flaw is there on a lot of players. There is a website wich I can´t find anymore wich has some audioexamples af a commercial (released) mix, wich you can listen to at -3dbf to keep your own converters out of the equation, with 1 the origanal mix (digtal clone) , 2 rerecorded through a converter that handels overs well and 3 a cd player (not a cheap one) that doesn´t handle overs well. The difference was very obvious!! there also was a delta file with just the distortion and that was realy shocking.
So remember it is not the distortion of the digital waveform but the clipping of the analog output stage that is the problem. But that can be serious clipping like 0.0001 at 0dbf and 10% at +1dbf