Audio Alchemy's EDR*S Processing Impressions
Daniel Baker
At the Stereophile Show in NYC this weekend, Audio Alchemy was
demonstrating their EDR*S remastering process. If you gave them a CD,
they would process a track or two for you and give you the CD-R (for
free) to try on your own system. If I am correct, EDR*S reads the
data of a CD, uses eight DTI Pro-32's to increase the resolution (I
believe to 24 bits; they split the audio band into seven segments and
process each segment with a single Pro-32, and then use the eighth to
sum), and then dithers down to 16 bits for transcription to a CD-R
that can be read by any CD player. I invite Mark Schifter of Audio
Alchemy to correct or elaborate on my description of the process, if
necessary.
The results? Pretty amazing. I was expecting an improvement, but not
an enhancement that was so *consonant with the music*. The soundstage
wasn't just larger, but each instrument came to occupy more of its own
space in the acoustic, so that there was a harmony of individual
musicians playing together rather than just sounds coming from
different places. Detail was improved, with instrumental decays
sounding more natural (*not* just more evident), timbres more
realistic with greater body (if appropriate), and greater dynamics.
Clarity of line was greatly enhanced -- it was much easier to follow
each instrument. Overall, quite a welcome surprise. And again, it
wasn't just like adding a reverb sound effect or re-equalization or
something like that. The improvements were uniformly realistic,
appopriate and complementary to the music.
A couple oddities, though all relatively insignificant and not
necessarily inherent to the process. First, on one track I had
processed, the channels seem reversed. Don't know why this would
happen, or why it only happened to one track. Second, there are a
couple "blips" on the transcription, perhaps due to the double speed
CD-R machine they were using. Third, my CD player sometimes has
trouble reading the discs when they are first loaded (then again, my
CD player -- Micromega Stage 3 -- *sucks* as far as disc toleration is
concerned). Other than the first problem, I don't think these really
have anything to do with the process itself.
A couple people I have spoken to have trouble with the concept of the
process as far as information theory is concerned (if you only have 16
bits on the regular CD, where are the new bits coming from?). I'm not
familiar enough with the theory behind this or the process itself to
draw conclusions here. Regardless, I find the changes to be
appropriate improvements, so it doesn't concern me terribly.
Also, I probably wouldn't get discs processed where I already know
they are good recordings. For example, Gabe's PGM recordings, along
with being expertly recorded using the best associated equipment, are
originally 20 or 24 bit and then carefully noise-shaped to 16 bits,
possibly in a similar manner to how EDR*S comes back to 16 bits after
the resolution enhancement. It seems EDR*S in this case might be
redundant here, and the 20 or 24 bit recordings in this case are true
original information, so no problem with information theory. Still,
the vast majority of recordings out there (especially older ones,
which seem to particularly benefit from the process) are not
audiophile quality.
So anyway, check this out! It's worth a listen. Contact Audio
Alchemy to find out who to call for more information (I have the guy's
card, but it's at home and I'm not).
Happy trails,
--Daniel Baker
Keith Allsop
Daniel,
There was a problem with the show equipment where the left and right
channels could get swapped - this will of course be fixed.
Also, did you get your disk remastered on Wednesday morning? The CD-R
was defective and many disks made on that day are defective. This is
the second time that the CD-R has broken - they do not seem to be very
reliable.
Finally, there have been reports in the press to the effect that disks
made on CD-R's occasionally have compatibility problems with regular
CD transports, although I have not come across this myself.
Your comments on the improvements due to EDR-S are very welcome, and
in line with other feedback we have received. Thanks.
Keith Allsop
Audio Alchemy
Werner Ogiers
Daniel Baker (ba...@cvrc.med.upenn.edu) wrote:
: that can be read by any CD player. I invite Mark Schifter of Audio
: Alchemy to correct or elaborate on my description of the process, if
: necessary.
I would also like to invite AA to do this.
You can't increase the resolution of a given recording.
You can copy it.
You can reduce its jitter (when it's jitter caused by irregularly
spaced pits on the CD) by copying the data to a pre-wobbled CDR-disc.
You can do whatever you want to make the CD sound nicer.
But you can't bring back or 'invent' what simply isn't there anymore.
------------------------------------------------------------------------------
Audio Visions -> http://www.esat.kuleuven.ac.be/~ogiers/welcome.html
------------------------------------------------------------------------------
Gabe M. Wiener
In article <4p89ma$n...@biosun.harvard.edu>,
Daniel Baker <ba...@cvrc.med.upenn.edu> wrote:
>A couple people I have spoken to have trouble with the concept of the
>process as far as information theory is concerned (if you only have 16
>bits on the regular CD, where are the new bits coming from?). I'm not
>familiar enough with the theory behind this or the process itself to
>draw conclusions here. Regardless, I find the changes to be
>appropriate improvements, so it doesn't concern me terribly.
It troubles me. But it is my job to be a skeptic until something is
proven to me to be worthwile.
I will preface the following by saying that while I have seen the
system and have spoken briefly to Keith Allsop (its creator), I have
not yet heard the system and have been hesitant to do so until I have
some questions answered.
Let us approach this question from the point of view of information
theory. In the realm of information theory, we have the concept of a
channel. A channel in this context is simply a pathway over which
information of any kind can be sent. Every channel has a certain
bandwidth, and while we needn't go into the whole question of what
defines a channel's bandwidth here, let it suffice to say that the
bandwidth regulates the amount of information that can be moved over
that channel in a specific amount of time.
Let us take the case of a 20-bit recording that I have made using a
20-bit A/D and a Nagra-D. That recording contains a certain amount of
information on it, and in fact the sheer quantity of information
present is far greater than the CD's ability to carry. As a result,
prior to my placing that information on the CD, I have to throw some
of that information away. Now, I can use all manner of neat-o tricks
to preserve as much of it as I can...to encode it into the bandwidth
that I have. I can use noise shaping or spread-spectrum hidden
bitstreams or any number of techniques. But in the end, I have
unequivocally reduced my channel bandwidth from 20 to 16 bits. What I
am left with is decidedly a 16-bit recording. I will never get those
4 bits back. Once they're gone, they're gone.
This is where my problem with the AA process comes in. No matter what
kind of algorithm it is, there is no way that any process can look at
those 16 bits of data and produce a different 16 bits of data that is
somehow closer to my original 20 bits. Think of it this way. If I
have a 5 decimal-place number, and I reduce it to 4 decimal places and
throw the fifth away, there is no way that anyone can look at those
four decimal places and deduce the fifth, and furthermore there is no
way for anyone to look at those four places and suddenly provide a
DIFFERENT four decimal-place number that is somehow "more correct."
No matter what you do, you cannot get a more accurate 16-bit number
from another 16-bit number. To do so would violate entropy, and this
was proven by Claude Shannon in his seminal paper on information
theory, "The Mathematical Theory of Communication," Bell System
Technical Journal, October 1948. Besides, it's common sense.
And thus I am left to wonder what the AA process is actually doing. I
have not yet heard it run on one of my recordings, but I would be
very, very surprised if the 16-bit processed disc sounded closer to
the original master tape than my 16-bit original disc. Once again,
without knowledge of what's on the master tape, there exists no
information from which they can discern data that isn't there. You
can't reverse entropy. There are a lot of DSP processes that one
could run on 16-bit audio to produce another 16-bit recording, but
as of this instant we have no reason to believe that the AA process
is taking us closer to the original.
When I visited Marc Schifter and Keith Allsop at Hi-Fi '96, they were
friendly but not wholly willing to reveal what it is that they're
actually doing. This is understandable given the high-end market, but
they have a ways to go if they want to gain the acceptance of their
process as a legitimate method of processing. Notice how all the
well-respected extant commercial processes (UV-22, SBM, Meridian, etc)
are published documents which you can read about. Even those who want
to be industrious can go read the HDCD patent application. Keith,
how about a paper at the next AES convention?
The algorithms need to be explained before I'll be convinced that it's
something I want done to my audio.
--
Gabe Wiener Dir., PGM Early Music Recordings |"I am terrified at the thought
A Div. of Quintessential Sound, Inc., New York | that so much hideous and bad
Recording-Mastering-Restoration (212) 586-4200 | music may be put on records
ga...@pgm.com http://www.pgm.com | forever."--Sir Arthur Sullivan
Bob Olhsson
In article <4pufs9$d...@decaxp.HARVARD.EDU>, "Gabe M. Wiener"
<ga...@pgm.com> wrote:
>...
>This is where my problem with the AA process comes in. No matter what
>kind of algorithm it is, there is no way that any process can look at
>those 16 bits of data and produce a different 16 bits of data that is
They don't claim to be creating 16 bits that are closer to your
original 20 bits. They are claiming to create 18 to 20 bits that may
be closer to your 20 bits than the a CD's 16 bits are. To turn their
20 bits back into 16 requires the same tricks that you used in the
first place to reduce 20 to 16.
My experience is that on some material (mostly older CDs) it is more
effective than other (mostly recent, real high quality CDs.) For sure
it IS signal processing and as such, it does make things sound
different, but then, that's the whole point. What amazes me is how I'm
generally not inclined to bypass it on newer recordings as I would
expect to be logically.
--
Bob Olhsson Audio | O tongue, thou art a treasure without end.
Box 555,Novato CA 94948 | And, O tongue, thou art also a disease
415.457.2620 | without remedy. == Jelal'uddin Rumi ==
415.456.1496 FAX |
Andre T. Yew
ogi...@imec.be (Werner Ogiers) writes:
>You can't increase the resolution of a given recording.
This is not strictly true. If you know the signal follows a
certain set of rules, you can theoretically reconstruct it. A bunch
of people in the 70s did this for pictures, and you can also ask some
math friends about analytic continuation.
--Andre
--
PGP public key available
TomMorley
ogi...@imec.be (Werner Ogiers) writes:
>You can't increase the resolution of a given recording.
But music is not random bits. The same is true of images. This is why
image compression can work. (see for instance Fractals Everywhere, by
Michael Barnsley , by the way I'm thanked in the preface!)
Tom Morley
(A Mathematician by Trade)
Gabe M. Wiener
In article <4q69bj$f...@tolstoy.lerc.nasa.gov>,
Andre T. Yew <and...@alumnae.caltech.edu> wrote:
> This is not strictly true. If you know the signal follows a
>certain set of rules, you can theoretically reconstruct it. A bunch
>of people in the 70s did this for pictures, and you can also ask some
>math friends about analytic continuation.
I still maintain that better results will be obtained if you start
with a higher-resolution master and use an intelligent approach to
reduce it to 16 bits, rather than using an algorithm (no matter how
good) to go from 16 bits to 16 bits.
James Durkin
tomm...@aol.com (TomMorley) writes:
> ogi...@imec.be (Werner Ogiers) writes:
>> You can't increase the resolution of a given recording.
> But music is not random bits. The same is true of images. This is
> why image compression can work.
Would you care to elaborate on this? This statement, in and of
itself, doesn't seem to say a heck of lot. Just because music isn't
random, doesn't imply that you know enough about the underlying
continuous signal, given a limited resolution approximation, to
reconstruct it at some higher resolution.
[[ James W. Durkin -- j...@graphics.cornell.edu ]]
[[ Program of Computer Graphics -- Cornell University ]]
Bob Myers
> >You can't increase the resolution of a given recording.
> But music is not random bits. The same is true of images. This is why
> image compression can work. (see for instance Fractals Everywhere, by
> Michael Barnsley , by the way I'm thanked in the preface!)
A better original statement would've been that you can't increase the
INFORMATION CONTENT of a given recording. 16 bits is 16 bits is 16
bits. You can use the available bandwidth more efficiently, but you
can never pack more than a certain amount of information in to a given
channel. Image compression, or for that matter ANY compression
scheme, works either by removing redundant information (which can be a
lossless process), or by throwing away information that isn't
redundant but which we don't think we care much about (which by
definition is a lossy process). But in no case can you get (a) more
information out of the process than was present in the original, or
(b) somehow get around the theoretical limits of the channel capacity.
To get back to the original point, there is simply no way you can pack
20 bits of honest-to-good useful information into 16-bit samples
without doing some form of compression which results in an average
LOSS of information. The result may or may not be better than what
you'd hear having gone straight from the analog signal to 16 bits in
the first place. I suppose the one advantage I can see in trying to
pack 20 into 16 is that in effect, you would be doing some dithering
which could be controlled a little better than otherwise.
Bob Myers KC0EW Hewlett-Packard Co. |Opinions expressed here are not
Workstations Systems Div.|those of my employer or any other
my...@fc.hp.com Fort Collins, Colorado |sentient life-form on this planet.
Alex Lee
James Durkin <j...@graphics.cornell.edu> writes:
>tomm...@aol.com (TomMorley) writes:
>> ogi...@imec.be (Werner Ogiers) writes:
>>> You can't increase the resolution of a given recording.
>> But music is not random bits. The same is true of images. This is
>> why image compression can work.
>itself, doesn't seem to say a heck of lot. Just because music isn't
>random, doesn't imply that you know enough about the underlying
>continuous signal, given a limited resolution approximation, to
>reconstruct it at some higher resolution.
I believe that he is talking about Fractal compression. From how it
has been explained to me, it takes the random data, and then tries to
find a fractal equation that closely APPROXIMATES the data. It only
compresses part of the picture in an given equation, of course, but
then the stored equation is much smaller then the data it represents.
When compressing the data,you specify the target compression ratio,
and the how much time the compression program should spend finding the
best fractal equations. It supposedly works very well if you give it
a lot of time, but as the more compressed and the less time you
spend,the more noticable artifacts from the compression become. You
can actually perceptually increase the quality of a bad picture, and
enlarge it a bit without a problem.
The problem with this is that it is ONLY an APPROXIMATION of the data.
Another problem is that it is very proprietary,and the inventors,
seeing a good thing, charge large licensing fees.
Alexander Lee
zo...@panix.com
Bob Olhsson
In article <4q9qv9$n...@agate.berkeley.edu>, James Durkin
<j...@graphics.cornell.edu> wrote:
>tomm...@aol.com (TomMorley) writes:
>> ogi...@imec.be (Werner Ogiers) writes:
>>> You can't increase the resolution of a given recording.
>> But music is not random bits. The same is true of images. This is
>> why image compression can work.
> Would you care to elaborate on this? This statement, in and of
> itself, doesn't seem to say a heck of lot. Just because music isn't
> random, doesn't imply that you know enough about the underlying
> continuous signal, given a limited resolution approximation, to
> reconstruct it at some higher resolution.
I would think that truncation might well follow enough of a pattern to
be at least somewhat "patchable."
It's still really got to be a crap-shoot whether the result is more or
less accurate, however, the effect IS pleasing, and a lot more like
that of having used a 20 bit converter instead of a 16, at least to my
ears.
Redithering is a new artform that certainly has promise at least for
older material.
We ought not to get overly hung up on accuracy. Certainly it is a
useful tool and it IS very important to remain conscious of the
difference between accuracy and an inaccurate but pleasing enhancement
effect.
One must always remember that pleasing inaccurate enhancements are the
name of the commercial recording game. Nobody ever sits in the typical
mike's catbird seat yet no one would consider requiring that mikes
only be used between 3 and 6 feet off the ground in an audience
seat. Likewise some of the most famous classical recording venues in
the world are routinely "sweetened" with artificial reverb. Nobody
pays much attention because the recordings made there WORK for their
audience which is the only truly valid criteria. When an effect
distracts because of heavy-handed use is when it becomes
objectionable.
qaq...@biostat.sph.unc.edu
In <4q69bj$f...@tolstoy.lerc.nasa.gov>, and...@alumnae.caltech.edu
(Andre T. Yew) writes:
>ogi...@imec.be (Werner Ogiers) writes:
>>You can't increase the resolution of a given recording.
> This is not strictly true. If you know the signal follows a
>certain set of rules, you can theoretically reconstruct it. A bunch
>of people in the 70s did this for pictures, and you can also ask some
>math friends about analytic continuation.
1. Would you care to give us one example from that "set of rules"?
2. Let us assume that we know that set of rules. How can this
knowledge be used to convert one 16-bit stream into another 16-bit
stream that is closer (in some sense) to the original signal even
without knowing how the original signal was transformed into a 16-bit
signal?
3. I know what people do with pictures (image estimation /
reconstruction / processing) and it bears no resemblance to what is
being discussed here.
bq
TomMorley
Not to get into this but here's my .02
>>You can't increase the resolution of a given recording.
>1. Would you care to give us one example from that "set
> of rules"?
No I, but surely bits 17 through 20 as "reconstucted" are not
compleatly arbitrary. You may have lost Forever the 2dn violinist in
the third row, fourth over from the right shuffing his feet, but the
ambience from the flute that just dropped below bit 16 can (in
principle) be guessed at.
>2. Let us assume that we know that set of rules. How can
>this knowledge be used to convert one 16-bit stream into
>another 16-bit stream that is closer (in some sense) to
>the original signal even without knowing how the original
>signal was transformed into a
>16-bit signal?
I agree what can 16 to 16 do? One possible answer is that this would
take advantage of know properties of D to A converters. This smacks of
the various failures in the vinyl? (Help be out here RCA?? late
60's???)
>3. I know what people do with pictures (image
>estimation / reconstruction / processing) and it bears
>no resemblance to what is being discussed here.
But 16 to 20 bit might be analogous to image restoration or
enhancement.
General comment: A great deal of the CD's on the shelf next to this
computer are acoustic blues from the 20's and 30's. The CD are taken
from old rare 78's often in wretched condition. Noetheless, dispite
the above comments, I find most current attempts at retoration (such
as CEDAR) unsatisfying and unmusical, especially when used with a
heavy hand. I usually prefer (at this point in technology) a straight
78 to CD transfer, finding the best of the (say) 4 known copies of the
78.
There is something vaguely perverse about listning to Charlie Patton's
High Water Everywhere (part II) through state of the art (inso far as
my budget will permit) equiptment.
Happy Listening!
Tom Morley
Tim Takahashi
TomMorley <tomm...@aol.com> wrote:
>General comment: A great deal of the CD's on the shelf next to this
>computer are acoustic blues from the 20's and 30's. The CD are taken
>from old rare 78's often in wretched condition. Noetheless, dispite
>the above comments, I find most current attempts at retoration (such
>as CEDAR) unsatisfying and unmusical, especially when used with a
>heavy hand. I usually prefer (at this point in technology) a straight
>78 to CD transfer, finding the best of the (say) 4 known copies of the
>78.
I too have been seriously bitten by the 78rpm bug, and more recently
old, old, old R&B.
My introudction into the world of tube amps came at the beginning of
my 78rpm collecting kick (about 10 years ago). I had dug out my dad's
old early 1950's Bogen mono-integrated-tube-amp for use in playing
back 78rpm disks. This amp had adjustable equalization for both treble
rolloff and bass turnover frequencies.
A correct playback equalization is essential when playing archival
material. Even Columbia "LP" pressings from the pre-RIAA days benefit
greatly.
In any case, experience with the lush sound of a classic 6L6 push-pull
tube amp got me going on the high-end kick. Home-brew 12AX7A pre-amp
and slightly modified Dyna ST-70 later; the Magnepans are happy as are
the 78s.
Speaking of 78->CD transfers, have you actually spun 78s? Charley
Patton is rare stuff.... but the sound of even a worn 78 is alluring.
I've never heard a remastering that fully captures the "sound" of a 78
(neither cassette nor 7.5ips open reel does justice). A 78 seems to
have a very dynamic sound, despite the poor s/n ratio (though clean
78s are nearly as quiet as an LP).
The simple, uncompressed microphone feed used on many 78rpm era
recordings also contributes to the unmistakable liveness.
Some of the most "78"-like remasterings I've heard are those on the
Russian Melodiya label. Alas, they did not reissue much in the way of
Blues legends 8^).
>There is something vaguely perverse about listning to Charlie Patton's
>High Water Everywhere (part II) through state of the art (inso far as
>my budget will permit) equiptment.
probably more typical than you might imagine.
-tim
Gabe M. Wiener
In article <4q3o61$e...@tolstoy.lerc.nasa.gov>,
Bob Olhsson <o...@hyperback.com> wrote:
>They don't claim to be creating 16 bits that are closer to your
>original 20 bits. They are claiming to create 18 to 20 bits that may
>be closer to your 20 bits than the a CD's 16 bits are.
And they have not yet explained to us how this is possible. It flies
in the face of the most basic concepts of entropy.
>To turn their
>20 bits back into 16 requires the same tricks that you used in the
>first place to reduce 20 to 16.
And the net result is that they are taking a 16-bit dataset and
attempting to supersede it with another 16-bit dataset. This strikes
me as a little bit of a stretch, to put it euphemistically. If the AA
guys would like to give a cogent technical explanation of what's going
on, I'm all ears.
One can go out and read technical papers (AES or otherwise) on nearly
every commercial encoding process around.....be it UV-22, SBM,
Meridian, what have you.
When manufacturers refuse to explain what's up, that's when I begin to
get concerned.
Gabe M. Wiener
In article <4ppf6n$1...@eyrie.graphics.cornell.edu>,
Werner Ogiers <ogi...@imec.be> wrote:
>You can reduce its jitter (when it's jitter caused by irregularly
>spaced pits on the CD) by copying the data to a pre-wobbled CDR-disc.
Let's be correct in our terminology here. You can re-make its pit
geometry by copying it to a CD-R. Whether this re-making results in
less jitter on decoding...and whether this has any effect on sound
quality...is entirely a function of what hardware you use to play it
back. Geometrical differences are not ipso facto translated into
jitter, but evidence suggests that they can be under certain
circumstances.
>But you can't bring back or 'invent' what simply isn't there anymore.
Well, one could "invent" whatever one wants. But in no way can you
recover that which has been discarded. You can use whatever
heuristics you wish in order to synthesize or interpolate data, but
that does not in any way suggest that this data resembles the
original.
Let us not forget the most fundamental rule of entropy. Once you have
lost information entropically (i.e. by collapsing it), you cannot get
it back.
Mark Schifter
>When manufacturers refuse to explain what's up, that's when I begin to
>get concerned.
Gabe...
Our company holds some 14 patents (via Peter Madnick and others)...
and YOU of all people should be "smart enough" to realize that once
these sorts of things are published they become "road maps" for others
to follow (read steal)...
I'm going to ask YOU to sign a non-disclosure document and then ask
Keith to visit you and make some re-mastered copies (and explain the
process) for you. I'll be curious as all get up to hear about your
reaction.
We've made several hundred re-mastered copies for truth seekers like
yourself (self proclaimed or otherwise)... and EVERYONE we've made
copies for (of those that have reported) have been startled by the
outcome. This IS a fact. So it's probably time to explain things a bit
further to one such as yourself.
Gabe M. Wiener
In article <4qjtvp$1...@agate.berkeley.edu>,
Mark Schifter <LWB...@prodigy.com> wrote:
>Our company holds some 14 patents (via Peter Madnick and others)...
>and YOU of all people should be "smart enough" to realize that once
>these sorts of things are published they become "road maps" for others
>to follow (read steal)...
I am well aware of the downside of patenting one's technology. But I
am also aware of the fact that without a frank and open dialogue with
the audio engineering community, it is a difficult thing for those of
us who use the technology to consider its use for professional
purposes.
>I'm going to ask YOU to sign a non-disclosure document and then ask
>Keith to visit you and make some re-mastered copies (and explain the
>process) for you. I'll be curious as all get up to hear about your
>reaction.
If I am presented with a non-disclosure document, I will sign it and
will not discuss the nature of the algorithms or the functioning of
the technology itself.
But as I'm sure you can appreciate, what I will do is report quite
honestly and fairly what my subjective findings were, after evaluating
the technology and the effect that it has on my software.
I consider myself one of the fairest testers I know. I have no
loyalty to anyone except he who makes the best products. Every so
often I come across a product line (like Prism, for instance) who puts
out consistently superior equipment. But even there, I always say to
Graham (director of Prism), "Be careful when you ask me to run a
shootout of your converter against someone else's, because you have to
be prepared for the possibility that you might lose."
Of course, in 2+ years, they haven't yet, but my point is that if I
feel a technology is good, I will say so, and if I feel it isn't, I'll
say so too.
I await the non-disclosure document. After I sign it and return it,
please ask Keith to call me at work to make an appointment to visit.
And, folks, there we leave this issue until after Keith's visit.
Gabe M. Wiener
In article <4qcjkb$f...@eyrie.graphics.cornell.edu>, Bob Olhsson
<o...@hyperback.com> wrote:
> It's still really got to be a crap-shoot whether the result is more
> or less accurate, however, the effect IS pleasing, and a lot more
> like that of having used a 20 bit converter instead of a 16, at
> least to my ears.
If it's a crap-shoot, then I can assure you that the odds are weighted
toward the less accurate.
I have many DSP units that take in 16-bit inputs, do a little
processing (whatever algorithm might be running) and output 20 or 24
bits. But that does not mean that I have any more information about
my original recording. It means that the DSP algorithm just so
happens to have a higher precision output than my original signal!
> We ought not to get overly hung up on accuracy. Certainly it is a
> useful tool and it IS very important to remain conscious of the
> difference between accuracy and an inaccurate but pleasing
> enhancement effect.
No, we ought to get exceedingly hung up on accuracy. The goal of
audio equipment...at least in my work...is reproducing the musical
experience, not editorializing on it. If the performance is warm and
voluptuous, or cold and stark, I want that rendered on the CD.
> One must always remember that pleasing inaccurate enhancements are
> the name of the commercial recording game.
Au contraire. Many (myself included) strive to add as little
coloration to the recording as possible, and to let the music speak
for itself.
> Nobody ever sits in the typical mike's catbird seat yet no one would
> consider requiring that mikes only be used between 3 and 6 feet off
> the ground in an audience seat.
Well, excepting the fact that the binaural folks do this all the time,
let us remember that we place microphones the way we do in order to
compensate for the fact that the recording still has to be reproduced
through loudspeakers, which means that the sound gets radiated twice.
Once from performer to microphone, and once from loudspeaker to ear.
Contrast this with the single radiation that occurs for a live
performance. If everyone recorded binaurally, placing the mics six
feet up in an audience seat would be precisely what we'd do.
> Likewise some of the most famous classical recording venues in the
> world are routinely "sweetened" with artificial reverb.
I for one would never be caught dead in a venue that needed reverb
unless someone put a gun to my head.
> Nobody pays much attention because the recordings made there WORK
> for their audience which is the only truly valid criteria.
Do they? Most such recordings, in my experience, sound appalling, and
nothing like real music in a real hall. Artificial reverberation and
excessive processing may be fine for pop music, but as far as I'm
concerned it has no place in classical recording, or any recording
genre where the goal is not to create sounds but to re-create a live
musical event.
Steve Zipser (Sunshine Stereo)
Gabe M. Wiener wrote:
> And the net result is that they are taking a 16-bit dataset and
> attempting to supersede it with another 16-bit dataset. This strikes
> me as a little bit of a stretch, to put it euphemistically. If the AA
> guys would like to give a cogent technical explanation of what's going
> on, I'm all ears.
Gabe:
Why don't you, one of the most refreshingly open-minded engineers I've
had the pleasure of meeting, just go listen to the damn thing first.
At least go listen to their DDTIPRO32. Then put it on the test bench
- as I'm sure you could measure what it does :)
> One can go out and read technical papers (AES or otherwise) on nearly
> every commercial encoding process around.....be it UV-22, SBM,
> Meridian, what have you. When manufacturers refuse to explain what's
> up, that's when I begin to get concerned.
Perhaps they are waiting for a patent?
Zip
Bob Olhsson
In article <4qjoeb$p...@decaxp.HARVARD.EDU>, "Gabe M. Wiener" <ga...@pgm.com>
wrote:
> When manufacturers refuse to explain what's up, that's when I begin
> to get concerned.
They have explained that they are interpolating more bits
somehow. They only are not talking about what kind of algorhythm they
are doing it with.
I think AA is too small a company to be able to afford the several
years of secret development and non-disclosure agreements that
preceeded most of the AES papers and explainations you are referring
to.
The Apogee UV-22 and Meridian noise-shaping technology were both
derived from previous academic research. Neither company has told
anybody HOW they do what they do, only that they are accomplishing
what the academic research suggested might be do-able ten years
ago. There hasn't been a comparable public feasability discussion
preceeding the AA work however I can't see blaming that situation on
Audio Alchemy.
It would be interesting to compare a 20 bit recording with an
interpolated 20 bit result from the same recording reduced to 16
bits. Hopefully I can get some time to try it.
jj, curmudgeon and all-around grouch
In article <4q9qv9$n...@agate.berkeley.edu> James Durkin
<j...@graphics.cornell.edu> writes:
> Would you care to elaborate on this? This statement, in and of
> itself, doesn't seem to say a heck of lot. Just because music isn't
> random, doesn't imply that you know enough about the underlying
> continuous signal, given a limited resolution approximation, to
> reconstruct it at some higher resolution.
You can't do that. Once you've added noise, you're stuck with it,
unless you have other knowledge of the exact signal structure.
But you can reduce the bit rate, by using source-coding techniques
(source modelling, rate-distortion coding, noiseless coding) to reduce
the bit rate while maintaining a 1:1 input to output.
It is conceivable that one could encode on a CD whatever resolution
(it would be time-varying, but not worse than 16 bits) could be
represented in 16 bits, by using some sort of very flexible source
coder.
There are, of course, better ways, one might also run a very-high-rate
perceptual coder, use that knowledge of perception that the high-end
proponents in this group adamantly reject, and get perhaps 20 bits of
input 'resolution' where it matters. (one could get more by
compression, but getting meaningful bits to the input might be a bit
tough)
A look in Akansu and Smith "Subband and Wavelet Transforms, Design and
Applications", Chapter 9, might unconfute some of the differences
between source coding gain, perceptual coding, and "what works".
--
Copyright akalice!jj 1996, all rights reserved, except transmission by USENET
and like facilities granted. This notice must be included. Any use by a
provider charging in any way for the IP represented in and by this article
and any inclusion in print or other media are specifically prohibited.
jj, curmudgeon and all-around grouch
In article <4q9r1k$n...@agate.berkeley.edu> my...@hpfcla.fc.hp.com (Bob
Myers) writes:
> To get back to the original point, there is simply no way you can pack
> 20 bits of honest-to-good useful information into 16-bit samples
> without doing some form of compression which results in an average
> LOSS of information.
Whoops. Gotta pick at a nit here, sorry.
If you have a full 20 bits of real, non-redundant information, you're
right.
But that's not what most audio signals look like. Even the
least-redundant audio signal I've ever measured (gets out Akansu and
Smith again), has about -30 dB give or take, spectral flatness
measure, indicating that a purely rate-distortion coding method can
remove about 5 bits on average from that signal, WITHOUT CHANGING THE
VALUES OF THE DECODED SIGNAL.
Now, these two statements are not contradictory, what that -30dB
spectral flatness measure shows is that the signal, while it might
have 16 or 20 bits input, has redundancy (i.e. each sample can be
predicted from the past samples) that can be extracted via well-known
methods and utilized for coding gain. (That gain at a 64 sample
interval, to be clear.)
In other words, the amount of "information" in a 20 bit signal is at
MOST 20 bits/sample. It's possible to have much much less.
> The result may or may not be better than what you'd hear having gone
> straight from the analog signal to 16 bits in the first place.
Well, if the signal is typical of audio signals, it's likely to be
possible to do a 1:1 20 bit encode into 16 bits for nearly all samples
of the signal, and very likely all samples given the averages observed
on most audio signals. (A signal that could not be compressed in this
way would be white noise. I'm not sure we care all that much, but I
must be clear, compressing "white noise" is an oxymoron.)
> I suppose the one advantage I can see in trying to pack 20 into 16
> is that in effect, you would be doing some dithering which could be
> controlled a little better than otherwise.
For most, if not all audio signals, you could probably pack the full
information in the 20 bit signal into 16 bits, in a real process,
that restores the same numbers at the output. While it's certainly
possible to make a signal for which this wouldn't work, it's not very common
in "real music" signals.
Andre T. Yew
qaq...@biostat.sph.unc.edu writes:
>1. Would you care to give us one example from that "set of rules"?
How about three? A signal that is bandwidth-limited, and
point- sampled can be completely reconstructed by a sinc function.
Both signals still have 16-bit resolution, but the reconstructed
signal comes closer to the original signal than the point-sampled
version, answering your question 2.
Knowing the inital trajectory and velocity of Voyager, and the
positions and velocity of the various planets in our solar system, and
applying a set of rules known as "Newtonian physics", we get to derive
the complete trajectory at every point of Voyager.
Knowing the starting point and algorithm of a pseudo-random
number generator, you can retrieve a message hidden deep in otherwise
innocuous noise sent to you by a friend encoded with the same PRN
generator. For examples, spread-spectrum communications or
encryption.
Have you asked a math friend what analytic continuation is?
If so, you get a fourth example for free!
>3. I know what people do with pictures (image estimation /
>reconstruction / processing) and it bears no resemblance to what is
>being discussed here.
No, it doesn't. However, if you reread my original post
carefully, you will see that that wasn't what I was trying to say. To
summarize for you, someone said that you can't end up with more bits
than you started with, in effect, no more information than before. I
said that this is not strictly true, if you assume that the signal
follows a certain set of rules. Conservation of information is still
going on here, since you have brought in extra information for the
reconstruction by "knowing" that the signal followed a certain model.
Whether that assumption is valid or not is the interesting part, no?
And shouldn't that be the real question that we should be asking Audio
Alchemy?
Werner Ogiers
Mark Schifter (LWB...@prodigy.com) wrote:
: Our company holds some 14 patents (via Peter Madnick and others)...
: these sorts of things are published they become "road maps" for others
: to follow (read steal)...
But patent texts are public, and so anyone persistent enough can get
access to your patents, which should contain enough information to
understand and implement the invention. One is even allowed to do so
(implement, that is) for non-commercial or academic use.
: I'm going to ask YOU to sign a non-disclosure document and then ask
What's the point here? If I wanted I could retrieve all your patents
and publish them on the net within 2 weeks time, and without having to
leave my desk. And it would be legal.
Any company that subsequently would bring out products based on those
patens, however, is bait for your attorneys. Well, provided said
patents hold.
--
Werner Ogiers IMEC, division MAP
-------------------------------------------------------------------------------
Audio Visions -> http://www.esat.kuleuven.ac.be/~ogiers/welcome.html
-------------------------------------------------------------------------------
jj, curmudgeon and all-around grouch
In article <4qc6u4$o...@agate.berkeley.edu> zo...@panix.com (Alex Lee)
writes:
>James Durkin <j...@graphics.cornell.edu> writes:
>> Would you care to elaborate on this? This statement, in and of
>> itself, doesn't seem to say a heck of lot. Just because music isn't
>> random, doesn't imply that you know enough about the underlying
>> continuous signal, given a limited resolution approximation, to
>> reconstruct it at some higher resolution.
There are two things going on here.
1) Some people are talking about REMOVING noise from a signal, and
"raising the resolution". Well, you might, if you try hard, perhaps
change the noise spectrum, but for any signal that is not completely
deterministic, you're NOT going to "remove noise", and what's more you
will NOT reproduce the effect in a 16 bit PCM DAC.
Now, there may (or may not) be something to be had in changing the
noise spectrum, but this is NOT equivelent to providing, for example,
20 bits of resolution out of a 16 bit CD player. Furthermore, even if
you have a 20 bit DAC, you're not going to get 20 bits out of
something originally encoded in 16 bit PCM.
2) Signals are generally NOT random. Furthermore, in the sense of
compression, PURELY RANDOM signals are not noiseless compressable. By
that I mean signals that have an autocorrelation spectrum that is
zero, except at zero, or alternately a frequency spectrum that is
flat.
Signals that are generated from markov processes are, on the other
hand, predictable to the extent that the markov process adds
redundancy. What's more, many physical processes take a random or
pitchy pseudo-random excitation and filter it through something that
behaves a lot like a markov process. Some examples of this sort of
process are speech generation, and the function of the individual pipe
in a pipe organ. Signals generated in this fashion may occupy many
more bits in PCM than they have, in a strict mathmatical sense,
'information', and by using something called a "source coder" they can
reduce the number of bits in the signal without being lossy. (Said
coders can also be instantiated in a way that is quite lossy, btw.) A
good reference for this sort of process is Jayant and Noll's book
"Digital Coding of Waveforms" unfortunately now out of print. (While
I know of several books in preparation, none are out yet, sorry!)
>I believe that he is talking about Fractal compression.
Um, no reason to believe that. Fractal compression is a particular
source-modelling approach that uses fractals as the "source model".
It's applicable to some particular problems in image coding, but isn't
so very applicable to music signals, or normally to speech coding,
either. Basically, fractal coding is useful in the case where there
are differently scaled self-similarities in the signal. This is often
the case in some kinds of images. It is not usually the case in music
or other audio signals in the waveform sense. (There is some argument
that one might fractally code tunes or some such, but that's on a
different wavelength entirely, as it were.)
>The problem with this is that it is ONLY an APPROXIMATION of the data.
All models are only an APPROXIMATION. Many models use the
approximation to reduce the signal energy (i.e. number of bits
required to send it) and then send the DIFFERENCE between the signal
and the approximation.
This can be true of fractal compression, although it's not usually
done. It is very true of DPCM, ADPCM, LPC and other predictive coders.
(Please see Jayant and Noll, I'm not about to describe my entire
speciality in one article. It took Jayant and Noll about 700 pages at
graduate level, and the book is quite old (although not moot), so the
book would be about 4x the size now.
Stewart Pinkerton
ga...@pgm.com (Gabe M. Wiener) writes:
>> Likewise some of the most famous classical recording venues in the
>> world are routinely "sweetened" with artificial reverb.
> I for one would never be caught dead in a venue that needed reverb
> unless someone put a gun to my head.
>> Nobody pays much attention because the recordings made there WORK
>> for their audience which is the only truly valid criteria.
> Do they? Most such recordings, in my experience, sound appalling, and
> nothing like real music in a real hall. Artificial reverberation and
> excessive processing may be fine for pop music, but as far as I'm
> concerned it has no place in classical recording, or any recording
> genre where the goal is not to create sounds but to re-create a live
> musical event.
As it happens, I used to run a group of consultants which included
Phil Walsh, who was largely responsible for the asssisted resonance
system in the Royal Festival Hall, London. It's my understanding that
the system is not switched on when recordings are being made, although
I could certainly be wrong in the case of 'live' recordings of public
concerts.
--
Stewart Pinkerton | If you can't measure what you're making,
A S P Consulting | how do you know when you've got it made?
(44) 1509 880112 |
Matt Kennel
Mark Schifter (LWB...@prodigy.com) wrote:
:> When manufacturers refuse to explain what's up, that's when I begin to
:> get concerned.
: Our company holds some 14 patents (via Peter Madnick and others)...
: and YOU of all people should be "smart enough" to realize that once
: these sorts of things are published they become "road maps" for others
: to follow (read steal)...
If you have the patents then there is no reason to fear publication.
That is entirely the point of patents.
What are the patent numbers? The abstracts are available on line for
free.
--
Matthew B. Kennel/m...@caffeine.engr.utk.edu/I do not speak for ORNL, DOE or UT
Oak Ridge National Laboratory/University of Tennessee, Knoxville, TN USA/
Mark Schifter
> I await the non-disclosure document. After I sign it and return it,
> please ask Keith to call me at work to make an appointment to visit.
>
> And, folks, there we leave this issue until after Keith's visit.
We shall do so post haste...
-
MARK SCHIFTER
LWB...@prodigy.com
mobius
>Bob Olhsson <o...@hyperback.com> wrote:
>
>>They don't claim to be creating 16 bits that are closer to your
>>original 20 bits. They are claiming to create 18 to 20 bits that may
>>be closer to your 20 bits than the a CD's 16 bits are.
>
>And they have not yet explained to us how this is possible. It flies
>in the face of the most basic concepts of entropy.
>
>>To turn their
>>20 bits back into 16 requires the same tricks that you used in the
>>first place to reduce 20 to 16.
>
>attempting to supersede it with another 16-bit dataset. This strikes
>me as a little bit of a stretch, to put it euphemistically. If the AA
>guys would like to give a cogent technical explanation of what's going
>on, I'm all ears.
I haven't heard one of these discs yet, but I don't see a problem with
the notion of deriving a "better" 16-bit image; it hardly flies "in
the face of the most basic concepts of entropy." The key is that the
information on the CD is not random -- if it were, I would agree that
improvement would be an impossibility -- it's of course extremely
non-random. Since most text is non-random, it's not too difficult for
an automatic spell-checker to fix obvious typing errors. Perhaps a
closer analogy might involve improving graphic images on computer
displays while maintaining the resolution; without adding extra bits
or needing additional pixels, anti-aliasing software can mitigate
effects like curved or diagonal lines turning into stairsteps on
square pixels.
Of course a process like AA's wouldn't work if all CDs were,
given the constraints of the medium, perfectly mastered; although the
CDs I buy these days sound generally far better than ones produced ten
years ago, I wouldn't be surprised if there's still room for
improvement in the mastering process.
Alex Lee
In <4qoq77$8...@eyrie.graphics.cornell.edu> j...@research.att.com writes:
>In article <4qc6u4$o...@agate.berkeley.edu> zo...@panix.com (Alex Lee)
> writes:
>>I believe that he is talking about Fractal compression.
>Um, no reason to believe that.
Well now, if he wasn't I'll just retract that.
>>The problem with this is that it is ONLY an APPROXIMATION of the data.
>All models are only an APPROXIMATION. Many models use the
>approximation to reduce the signal energy (i.e. number of bits
>required to send it) and then send the DIFFERENCE between the signal
>and the approximation.
To be more specific about what I ment... When you digitize the sound
with the A/D, it is an approximation of the sound. When you compress
the data with a lossy compression algorithom, it is now an
approximation of the data which is in turn an approximation of the
sound. When you use a lossless compression algorithom, it should,
when uncompressed, be an identical copy of the data, which is an
approximation of the sound. I ment to peg fractal compression into
the former catagory, not the latter.
Now I realize that anytime you convert from a 20 bit master to a 16
bit recording, you will have to compress, and unless a new method of
compressing arises, it will most probably be lossy. That issue though
was beyond the intended scope of my previous post.
Alexander Lee
zo...@panix.com
D. S. Ritter
In article <4qmee1$3...@eyrie.graphics.cornell.edu>, j...@research.att.com
(jj, curmudgeon and all-around grouch) writes:
> For most, if not all audio signals, you could probably pack the full
> information in the 20 bit signal into 16 bits, in a real process,
> that restores the same numbers at the output. While it's certainly
> possible to make a signal for which this wouldn't work, it's not
> very common in "real music" signals.
And this reminds me of a compression scheme whose name eludes me. The
method was this: each data sample coming through was *not* the actual
absolute value of the signal, but the signed offset from the last
sample. Thus, a 16-bit sample size means that the maximum difference
from the *last* sample is 2^16 levels. There is no particular maximum
or minimum to the transition which can be accomplished within multiple
samples, and a 'return-to-zero' code makes transitions from the
extrema easy.
If I recall correctly, this method was (is?) used in the Micro$oft
WAVE format as one of the allowable compressions. Not that I consider
that much of an endorsement, but it is feasible, and easily
implemented in very fast hardware.
--
d...@spectra.net - this posting is *not* an official Spectra.Net policy.
D.S. Ritter 607 798 7300
"We did that, (well almost that), but it didn't work." - someone else
Bob Myers
Alex Lee (zo...@panix.com) wrote:
> To be more specific about what I ment... When you digitize the sound
> with the A/D, it is an approximation of the sound. When you compress
You don't digitize sound with an A/D. You digitize an "analog"
electrical signal, and that signal itself is ALREADY an approximation
of the sound. Given a sufficiently high sampling rate to cover the
bandwidth of interest, and enough bits so as to provide a dynamic
range exceeding that of the analog signal, nothing whatsoever is lost
in this conversion except errors introduced by inaccuracy in the A/D
circuit or its clock. But then, you add noise and distortion with
every analog stage a signal passes through as well, so there's no real
distinction at this point regarding the accuracy of the two encoding
methods.
> Now I realize that anytime you convert from a 20 bit master to a 16
> bit recording, you will have to compress, and unless a new method of
> compressing arises, it will most probably be lossy. That issue though
Why "most probably be lossy"? Lossless compressions exists - they
just generally aren't capable of nearly as high a compression ratio as
lossy schemes.
Robert Kay
but I figured you'd have questions, and I'm busy today and have to
post when I can. I hope the two messages arrive in the proper
order...
OK, let me make it a little more clear. There are two different
examples I'll show you to highlight the two different behaviors;
dither and over-sampling.
First we'll look at what dither can add:
Let's assume for a moment that the real signal we a sending is just DC
(i.e. 0Hz). This "signal" can be any level, but we know that it will
be quantitized ("binned", "integer-ized") at some point in the
process. Without dither this would in fact mean that we have forever
lost information about the level of the signal (because of its
extremely stationary nature); assuming it was between or (even worse)
well below the value of one of the "bits" on the a/d. There's nothing
we can do, because (within the stability of the system) the digital
output would be a constant stream of whatever integer number (zero if
it is too low).
Now, with the addition of dither (take to mean triangular distribution
for now) the low end bits are "twiddled" in a random way so they shift
between several values in a way that is proportional to the "true"
level of the signal. As I mentioned in the previous posting, to
"increase the resolution" of this signal we could fit a basis function
and then "re-synthesize" the signal at whatever bit depth we wanted
(although there is a limit to what accuracy is really available, it is
certainly more than 16 bits). In our test case above, with dc, it
only makes sense to use a "line" as a basis function. You can reduce
the fitting parameters of this to just finding the average of the
incoming signal. Let's say we buffer 32,768 samples. With our
dithered incoming signal being averaged we are able to "average-out"
much of the effect of the dither noise (assuming it was high-quality
dither noise in the first place that had a mean approaching zero) and
we are left with a number which more accurately represents the
original (theoretically infinite precision) signal. Once we have this
number we can spit out samples of higher bit depths, even adding
dither at a lower bit if you want.
I could go into much more detail about what the real limitations are
and give you a bunch of plots, but you get the idea. (I hope)
Now lets look at "over-sampling":
This is a little harder to explain, so have some patience as I
struggle through a way to do it, without pictures. Also, remember
that in real life you would also be including the effects of dither as
described above. I've purposely left it out so I could explain it
better.
We'll assume in this case that our signal is a 500Hz sine wave of some
level. Again, we'll digitize it at sixteen bits. (remember, no dither
this time, to make it easier). Again, the signal becomes
"integer-ized" at the sample levels.
Rarely in real life would any incoming signal have each crest and
trough sampled at the exact same moment in time relative to the zero
crossing of the sine wave. This could only happen if the sine wave
were in "perfect synch" with the sample rate. So, for each cycle we're
able to fill in additional "time points" of the sine wave. The
information is still only available at 16-bit resolution, but because
of the above, we get timing information filling in at lots of
different times over consecutive waves (this is not constrained by the
sample rate timing i.e., the phase at which the sine wave is advancing
relative to the sample rate might be very small)
How do we get more resolution ? Fit a basis function, then
re-synthesize again. In this case, we'll fit a sine wave ('cause we
know what we started with). With a big buffer of data, we have a lot
of points to work with when "fitting". Each one pushing and tugging
the level (and phase) for a "best fit. You end up with some simple
data on the output which describes the level and phase (or, real and
imaginary parts) of the sine wave.
Armed with these numbers we can re-synthesize a higher bit depth
signal.
Ok, great, so what ? You can make this work for dc or a sine wave;
but what about *music* ? Well, as I'm sure many of you know, music
can be decomposed into just these kinds of components.
This description is certainly an over-simplification; I still need to
be able to get work :-)
Of all of the "devil in the details" issues, the biggest problem is
knowing how to fit the basis functions correctly: Knowing what data at
what frequencies is likely to be original signal and which are likely
to be "noise" You don't want to fit to the noise or you'll just end up
with a perfect copy of your 16-bit original with some more bits tacked
on. The other big issue is what are the best basis functions for
music. Sine waves ? Some kind of wavelet function ? I'm not telling
...
robert kay
Robert Kay
here.
One of the most important is that a signal only "needs" to be sampled
at twice it's highest frequency (plus a little slop to make up for not
having a true "brick-wall" filter). Any "oversampling" will provide
some additional information. This is how "1-bit" converters (which
aren't really one bit) and the early cd players with only 14 bit d/a's
got all of the 16 bit signal out of less bits.
You can easily see that "most" musical information is at lower
frequencies and is therefore "oversampled" quite a bit with the
44.1kHz sample rate. This is an over-simplification of the situation,
but you can pull some extra information out of those 44.1kHz, 16-bit
samples. Add to this the fact that most modern recordings are
dithered, and there's plenty of info "buried" in there.
How do you get it out ? Well one way would be to fit a big bunch of
"basis functions" to the incoming 16-bit data; try to figure out
what's quantization noise and not dither in the fit functions (the
hard part), and then resynthesize at whatever bit depth you might want
to use. You might have to look at a hunk of data, but, what the heck,
the cd was already delayed from the time of recording to the time you
play it; what another few hundred milliseconds ?
What basis functions, well, you could use sine waves, or maybe
wavlets, there are lots of possibilites. The choice becomes almost an
art.
Are you "really" getting more information out of the 16-bit data ?
Well, you're not exactly, you're just using whats there better; using
the low-frequency "oversampling" inherent in the system along with
tendency of proper dither to average to zero. Will you every "add
output resolution" at 20kHz ? I don't think so; at 1000Hz, you bet !
I could easily give you a set of equations which would evaluate to a
curve of "infinite" precision, one which could never be completely
sampled at any bit depth. These equations can be (to some degree)
derived from the (properly dithered) samples on the original
recording.
It's not really all that hard,
robert kay
Curtis Leeds
> ...And this reminds me of a compression scheme whose name eludes me.
> The method was this: each data sample coming through was *not* the
> actual absolute value of the signal, but the signed offset from the
> last sample. Thus, a 16-bit sample size means that the maximum
> difference from the *last* sample is 2^16 levels. There is no
> particular maximum or minimum to the transition which can be
> accomplished within multiple samples, and a 'return-to-zero' code
> makes transitions from the extrema easy.
I believe you are referring to a digital system devised and proposed
by dbx. I think that they called it "delta modulation", and this was
before the CD era. My understanding was that it had a lot of promise,
but was rejected by Japan, Inc because of the "NIH" (not invented
here) syndrome.
Alex Lee
>You don't digitize sound with an A/D. You digitize an "analog"
>electrical signal, and that signal itself is ALREADY an approximation
>of the sound. Given a sufficiently high sampling rate to cover the
>bandwidth of interest, and enough bits so as to provide a dynamic
>range exceeding that of the analog signal, nothing whatsoever is lost
>in this conversion except errors introduced by inaccuracy in the A/D
>circuit or its clock. But then, you add noise and distortion with
>every analog stage a signal passes through as well, so there's no real
>distinction at this point regarding the accuracy of the two encoding
>methods.
I don't disagree with you. In order for brevity, when I said that the
process of digitizing sound is gives you a dataset that approximates
the sound, I ment the entire process, from miking to spitting bits out
to a digital recorder of some fashion.
>Why "most probably be lossy"? Lossless compressions exists - they
>just generally aren't capable of nearly as high a compression ratio as
>lossy schemes.
The second sentance answeres why I said that it would be most probalby
lossy. If the audio industry could compress 20 bits of data into 16
bits in a feasable (and all that implies for production of mass
consumer audio items) realtime manner, this discussion would probably
not be happening.
Alexander Lee
zo...@panix.com
Andre T. Yew
>With our
>dithered incoming signal being averaged we are able to "average-out"
>much of the effect of the dither noise (assuming it was high-quality
>dither noise in the first place that had a mean approaching zero) and
>we are left with a number which more accurately represents the
>original (theoretically infinite precision) signal.
Interesting thought. This is equivalent to applying a low-
pass filter to the data, which is not always desirable. However, what
if the low-pass filter were applied selectively, such as when only the
lowest 4 bits or so are active? Measurements I've seen in magazines
might imply that the DTI-Pro works like this. This would also explain
why the DTI-Pro is limited to lower-frequencies to do its work --- to
apply the low-pass filter, it would have to apply it over a long
series of samples and it might actually need a large number of samples
to do its selection analysis. This might also explain how a static
(ie., non-realtime) system like the EDR-S would do a better job when
it could load entire datasets and examine it at its own leisure and
perform non-causal analysis.
Why activate the filter only at low signal levels? The
quantization error (assuming incompetant mastering engineers were at
hand) is worse at lower signal levels.
>We'll assume in this case that our signal is a 500Hz sine wave of some
>level. Again, we'll digitize it at sixteen bits. (remember, no dither
>this time, to make it easier). Again, the signal becomes
>"integer-ized" at the sample levels.
>Rarely in real life would any incoming signal have each crest and
>trough sampled at the exact same moment in time relative to the zero
>crossing of the sine wave. This could only happen if the sine wave
>were in "perfect synch" with the sample rate. So, for each cycle we're
>able to fill in additional "time points" of the sine wave.
With all due respect, I disagree with this. The signal's
phase does not have to be the same as the sampling clock's phase to
perfectly reconstruct a signal that meets the Nyquist criteria.
>How do we get more resolution ? Fit a basis function, then
>re-synthesize again. In this case, we'll fit a sine wave ('cause we
>know what we started with). With a big buffer of data, we have a lot
>of points to work with when "fitting". Each one pushing and tugging
>the level (and phase) for a "best fit. You end up with some simple
>data on the output which describes the level and phase (or, real and
>imaginary parts) of the sine wave.
I disagree with this as well. Since most signals today are
point-sampled, the correct reconstruction function is a sinc function.
The sampled signal is a convolution of the signal, the low-pass
filter, and especially important, the sampling function.
Reconstruction with a sine signal would make you end up with only a
sine signal! The rest of the piano is gone.
qaq...@biostat.sph.unc.edu
[ quoted text deleted -- rgd ]
>having a true "brick-wall" filter). Any "oversampling" will provide
"brickwall filter"? Who uses it these days?
Oversampling is used not to provide "additional information" (the same
bandwidth is retained, hence OVERsampling) but to ease the filter design.
bq
Robert Kay
In article <4r1a5d$r...@agate.berkeley.edu>, and...@alumnae.caltech.edu
(Andre T. Yew) wrote:
> >Rarely in real life would any incoming signal have each crest and
> >trough sampled at the exact same moment in time relative to the zero
> >crossing of the sine wave. This could only happen if the sine wave
> >were in "perfect synch" with the sample rate. So, for each cycle we're
> >able to fill in additional "time points" of the sine wave.
>
> With all due respect, I disagree with this. The signal's
> phase does not have to be the same as the sampling clock's phase to
> perfectly reconstruct a signal that meets the Nyquist criteria.
I did not mean to imply that I thought that the *should* or *must* be
in synch to reconstruct the signal. My point was that in most cases,
for a sustained sine wave, the samples would occur at different times
relative to the sine wave itself. This is important because it does
provide more information about the signal. This would not be true if
the samples themselves were of infinite precision, but remember, they
are being quantitized at some finite level.
> >How do we get more resolution ? Fit a basis function, then
> >re-synthesize again. In this case, we'll fit a sine wave ('cause we
> >know what we started with). With a big buffer of data, we have a lot
> >of points to work with when "fitting". Each one pushing and tugging
> >the level (and phase) for a "best fit. You end up with some simple
> >data on the output which describes the level and phase (or, real and
> >imaginary parts) of the sine wave.
>
> I disagree with this as well. Since most signals today are
> point-sampled, the correct reconstruction function is a sinc function.
> The sampled signal is a convolution of the signal, the low-pass
> filter, and especially important, the sampling function.
Again, the fact that the signal has been "integer-ized" in the process
of sampling means that the perfect sinc function is not enough.
> Reconstruction with a sine signal would make you end up with only a
> sine signal! The rest of the piano is gone.
Of course, you do this with one sine wave and you end up with one sine
wave out. You do it with "all" sine wave frequencies and you get the
full signal. There are some tricks in here, but it works.
robert kay
jj, curmudgeon and all-around grouch
In article <4quint$8...@eyrie.graphics.cornell.edu> Curtis Leeds
<cle...@mail.idt.net> writes:
>I believe you are referring to a digital system devised and proposed
>by dbx.
The DPCM system described long predates dBx, it has been proposed for
use in speech coding for many years.
>I think that they called it "delta modulation", and this was before
>the CD era.
Well, the system predates the CD era by quite a bit (the first thing I
did for Bell Labs was build a realtime ADPCM coder, a more advanced
version of a DPCM coder), and it's not a delta modulator. A delta
modulator, or delta-sigma modulator, uses a 1-bit quantizer,
oversamples, and uses prediction and filtering to get an SNR advantage
over the 1-bit quantizer over (typically) the low-passed section of
the signal.
Ray Steele has written a very good book on delta mod and its variants.
Jayant and Noll, and later Bastiaan Kliejn, have written good books
that describe source-coding methods.
I must point out that these methods were pioneered by
speech-processing researchers, in the attempt to get more telephone
calls over the same channel.
>My understanding was that it had a lot of promise, but was rejected
>by Japan, Inc because of the "NIH" (not invented here) syndrome.
--
jj, curmudgeon and all-around grouch
In article <4qrd83$a...@eyrie.graphics.cornell.edu>
d...@tree.spectra.net (D. S. Ritter) writes:
>And this reminds me of a compression scheme whose name eludes me. The
>method was this: each data sample coming through was *not* the actual
>absolute value of the signal, but the signed offset from the last
>sample.
This is called "DPCM" for "differential pulse code modulation". It is
more or less the most primitive compression method in existance, and
assumes that there is a downward spectral slope of the signal. For
music, its use is "use with care" because of the need to choose an
implementation that does not introduce either slope limiting or
clipping.
Jayant and Noll's "Digital Coding of Waveforms" <title approximate,
I'm at home> has a detailed discussion of such algorithms, both of the
primitive type cited above, and of the more advanced versions.