posted 12 April 2002 18:52            

Here is a website that explains the phase shifting needed...only i am not that great at math.
http://hometown.aol.com/matrixquad/encode.htm

posted 12 April 2002 19:46            

I remember quadraphonic stuff from the early '70's. Is this the same thing?

posted 14 April 2002 12:08            

quote:

---

Originally posted by mathis:
during the nineties there was a renaissance in development of analog matrix encoding. 5-2-5 systems with stereo surrounds came to market (RSP). i have a preprint of 1997 written by david griesinger of lexicon, where he describes the progress in 5-2-5 matrix systems.

---

Hi Mathis,

Well just some listening experience from stereo to surround (2-2-7).

I use a Lexicon DC-1 surround pre-amp in the studio, the set-up in the studio are 4 loudspeakers (front and side, for the optimum result you must have a 7.1 system but my studio is to small for that), the decoding on the DC-1 (called Logic 7 for film and Music Surround for music, which is making use from the same technology of the Lares system) can restore the directional properties of recordings.

I do not know exact how this technique works, but if you listening to ordinary stereo music in surround the recreation of the original performance as well as the creation of a soundstage is amazing. This technique from Lexicon is going a bit farther then only delay and inverse steering, it is also using a ambience extraction method which provide very good results.

The interesting part of this is when you mute the front channels and listen to the side channels and hear what is used for surround it is mostly the amount of ambience, and the strange thing about this is; the DC-1 does not add anything to the sound!

posted 14 April 2002 16:52            

I think the Lexicon convert a quad LP very simple but what happens if you use the Soundfield microphone with the Lexicon? do you have 7.1 tracks! The fact is the DC-1 is a "consumer" product so this would be not ideal maybe the new generation the MC-12 put this in a pro-level with balanced in and outputs only one pessimistic thing it's very expensive compare this with a fully loaded Capybara so take a wild guess!...

posted 14 April 2002 16:58            

quote:

---

Originally posted by SSC:
"Have you tried listening to a stereo CD through the Kyma StereoToPseudoSurround module in the Prototypes? The "reverb" or "ambience" tends to end up in the surround speakers.

---

Not yet I'll try this out tomorrow.

Later,

posted 14 April 2002 18:57            

hi armand, to be honest, i canīt believe, lexicon dc-1 would help us for that application. on the quad lp itīs a specific analog computer which has to get decoded in its way ment by the developers. otherwise itīs something in the surrounds, but not what was initially ment.
Soudfield codes its information in a four discrete channel format, called B-format. there is three-dimensional information recorded.

the dc-1 does amazing results in seperating direct signals from ambient signals. it seems that the dc-1 can seperate dry signals from reverberant signals. i have no clue how this is possible from the algorithmical standpoint. any ides?

posted 15 April 2002 01:01            

BTW -- as for extracting ambience, some experiements here over the past month indicate that a good way to do this is to high-pass filter the sound to remove the most correlated frequencies in the bass region. Adding that high frequency stuff back in after some delay 1-20 ms enhances the soundstage in a manner you all describe. This experiment is very easy to perform with Kyma. Try a high pass filter with a cutoff around 500 Hz.

- DM

posted 16 April 2002 12:39            

Mathis, your right. I thought that Lexicon something mentioned about this in the manuel, but after reading the manuel again, they are saying following: Versions of the Atal/Schroeder/Damaske/Mellert technique have appeared in several consumer signal processors under various trade names and recordings such as those on Telarc. Lexicon incorporated these techniques. So here is my misconception.

David, I've tried this high-pass filter experiment put this in the surrounds and my impression is that the complete soundstage sounds a bit muddy.

posted 16 April 2002 13:17            

quote:

---

SSC, The StrereoToPseudoSurround module works fine only one thing somehow the low end frequencies are gone?-Armand[/B]

---

posted 16 April 2002 13:40            

quote:

---

Originally posted by armand:
David, I've tried this high-pass filter experiment put this in the surrounds and my impression is that the complete soundstage sounds a bit muddy.

-Armand

---

So, how does Lexicon and others extract the uncorrelated from the correlated without adding anything to the sound that wasn't already there, i.e., no artificial reverb nor added harmonics? Are you saying that Lexicon's technique accomplishes the expansion of soundstage without sounding muddy?

BTW, I'm curious what you actually did. "Muddy" infers bassiness, but the sound should have been passed through a high-pass filter. The resulting comb filtering should look a lot like a "room response" without the deep notches and peaks at the bass end. So how much delay did you incorporate on the foldback, what cutoff frequency did you use on the HPF, and did you attenuate the foldback audio? Did you use different delays on left and right? (e.g., 15 ms and 19 ms). Using differential delays causes a complex interleaving of the combs between left and right channels, also adding to the "room effects" simulation.

- DM

posted 16 April 2002 19:09            

HPFSurr.sit

The attachment gives a idea about this.

posted 16 April 2002 20:00            

As for 90 deg phase shift over all audio frequencies, it isn't really all that difficult, but you need the ability to create an FIR filter. You can also do it quite easily with an FFT. A 90 deg phase shift is also known as a Hilbert transform (actually the sum of the original signal and this quadrature signal becomes a Hilbert transform). For an odd number of taps in the FIR filter, the formula is simple: Set the middle tap to zero along with every other tap on either side. On the right, set the odd numbered taps (center being tap zero) to 1/N where N is the odd tap index. Do the same on the left, but make them negative. Voila! A Hilbert transformer. This has very broadband 90 degree phase shift. There will be droop at DC and at Nyquist, but otherwise it is all-pass with pure phase shift.

Some of my earlier submissions with FFT's also provided this using a slightly different technique. I think my "Sound Microscope" does this. Using a Hilbert transform (in FFT form here) allows us to create what is called an "Analytic Signal" which has a one-sided spectrum. This means that you can heterodyne with a complex exponential to shift the entire frequency axis anywhere you want. So, e.g., my sound microscope was used to narrow band filter a 100 Hz slice at 15 KHz in some old recordings and then slide this band down to audible range where I could listen for encoded information. (It turned out to be nothing more than a weak 15 KHz carrier, probably for tape bias purposes -- no encoded info!).

For a 90 degree phase shift under FFT's simply interchange the real and imaginary components, and negate the new real components. (Same as multiplying everything by i = Sqrt(-1). To go the other way, interchange the real and imaginary components and negate the new imaginary components (multiplying by -i = -Sqrt(-1)).

FFT's are probably the easiest way to go under Kyma. If you manage to implement a Hilbert FIR transformer, then you have to remember to delay the input signal by half the filter length before adding in quadrature. This is easy to do when the number of taps is odd. But if you use an even number of taps, then you have to delay by 1/2 sample period plus some integer number of periods. The FIR filter is also different in this case, but it doesn't suffer any droop at Nyquist - only at DC. Making a 1/2 period delay requires another kind of 1-pole all-pass filter.

- DM

[Note: I glossed over a lot of the details above... In an effort to ameliorate distortion and temporal aliasing artifacts one often multiplies these FIR Hilbert transformers with a Kaiser window to make them taper off a bit faster. Also, the longer the filter, the more accurate the phase shift over broad frequency ranges.

After re-reading Pete's submission above I see that he was almost performing the Hilbert transform with his FFT's. Without some kind of windowing applied to the outputs there will be time aliasing artifacts because the conjugation was applied effectively with a rectangular window (by default) and this rings quite a bit in time before and after the input signal. A waveshaper can be used inside the FFT Hilbert transformer to apodize this process. (Apodize: Greek = removing the feet -- same as "windowing" in FFT parlance).]

posted 17 April 2002 16:08            

When I read this, the first thing that struck me was that it sounded exactly like the formula for creating a square wave by adding together every other harmonic at amplitudes scaled by 1/N where N is the harmonic number. During lunch, we were having a heated discussion of whether or not that was just a coincidence when we suddenly came to an epiphany (right there in Cafe Kopi). It's because we *are* creating a square wave by adding up lots of odd numbered sine waves at 1/N amplitudes--but we're doing it in the frequency domain instead of the time domain! The result is a nice, half square wave that gives us the flat, all-pass response--i.e. no frequency is attenuated or boosted relative to any other.

In other words, think about the taps as delays, one sample apart from each other like these asterisks:

* * * * @ * * * *

with the @ being the center tap. According to David's formula, the coefficients on the first four are the negatives of the coefficients on the second four. So if you were to feed a signal, x, into the delay line and just take the taps on either side of the asterisk in isolation, you would have

x(5) - x(3)

which would be just like a comb filter (the signal added to itself delayed by two samples). If you look at the frequency response of this comb filter, you see that it cancels out at DC and the half sample rate and has its maximum in the middle. In fact, it is one half of a sine wave in the frequency domain.

Now take the next pair of taps outwards from the center. That gives you

x(6) - x(2)

This is another comb filter but its delay is 4 samples long, so its frequency response has an additional cancellation point in the middle of the spectrum (at 1/4 the SR). In fact, if you take phase into account, it looks like one cycle of a sine wave! And so on.

In other words, the frequency response of each pair of taps is a sine wave (in the frequency domain). Fourier says that you can create any signal by adding together enough sinusoids of the appropriate harmonic frequency and amplitude. So we should be able to create any frequency response by adding together enough pairs of taps with the appropriate delays and coefficients.

posted 17 April 2002 19:51            

- DM

[Ahh yes!! I see that you have given us a whole new Sound to play with -- FIRFilter. That opens up a whole universe of possibilities. Is it possible to dynamically alter the taps?

Also, I see that your Hilbert transform is Hamming windowed. I have always wondered about the use of the Hamming window. I knew that the raised cosine family of windows -- Hann, Bristow-Johnson, and even the Bartlett (triangular) window were perfect mixer windows.

I never looked closely at the Hamming window, but I find that it too is a perfect mixer, but with an amplitude of 1.08 instead of unity. So as one researcher to another, what are the other desirable properties of a Hamming window that makes one want to choose it over others? I know that it doesn't go all the way to zero at the edges. Is this a desirable feature? Or perhaps you choose it to maintain a narrower main lobe width than a Hann or B-J window?]

[This message has been edited by David McClain (edited 18 April 2002).]

Hmmm... I have been looking at the effects of these 90 deg phase shifts on sound and localization. A 90 deg phase shift can also be viewed as a frequency dependent time delay = 1/(4*f). This in turn means that the Haas effect is contributing to a sense of direction. But it also implies that you should hear an exaggerated localization at bass frequencies.

The Haas effect diminishes when the delay becomes near 20-40 microsec, but that would mean frequencies of around 10 KHz, so the apparent directionality will probably be apparent at all meaningful audio frequencies.

I have been curious about the purpose of adding a signal to its 90 deg phase shifted version. This seems to create a signal at 45 deg, or halfway between the mono mix in front and the L/R signals on the two sides. Unless additional delay is incorporated, I find it difficult to see how surround sound can be created from this simple addition. It would seem that all sounds are located in the hemisphere in front of the listener, extending only to extreme left and right sides, but never behind him/her.

What am I missing here?

- DM

posted 19 April 2002 16:09            

There is also a interview with David Griesinger on this link:
http://www.smr-home-theatre.org/lexicon/

Interesting fact (at this time) of what he says about discrete channels and matrix decoding.

David, I think you allready know it, this home page from David Griesinger:
http://world.std.com/%7Egriesngr/

There is a lot of info on technical papers.

-Armand

posted 20 April 2002 11:25            

Griesinger details quite nicely in this article
http://world.std.com/~griesngr/SUR.html

that matrix decoding systems do *NOT* present ideal decodings. I would say they can't -- simply because you can't extract correct 4 or 5 way sound from only two channels of information.

At best you can achieve some degree of separation for pseudo-surround by using various projection operators (the matrix). But he shows quite nicely how these systems appear to work, what their shortcomings are, and so on.

I see he is a physicist too! Very good!

Cheers,

- DM

Here is a very interesting statement from Griesinger...

"...We know from the physics of sound perception that it is possible to distinguish front sound from rear sound in two ways.

At low frequencies the two can be distinguished through small movements of the listener’s head. Small head movements cause predictable changes in the ITDs. Unfortunately head movements produce no shifts in ITDs when the sound field is largely diffuse.

The primary front/back cue at high frequencies is spectral. Sound that comes from the front of a listener has notches in the frequency response at about 8kHz. As a source moves from the front to the side and then the rear, these notches disappear around +-90 degrees, to be replaced by notches at about 5kHz when the source reaches about +-150 degrees. The solo violin spectrum is particularly rich in sound energy in the region of these directionally dependent spectral anomalies."

This spectal shaping accompanying panning should be very easy to accomplish with Kyma. It also indicates why I have trouble localizing high frequency sounds, given my hearing loss. These statements were found in the paper at:
http://world.std.com/~griesngr/objmeas.html

posted 20 April 2002 18:40            

The impression he gives would lead one to believe that fancy reverbs done with convolution of impulse responses are likely to yield unsatisfactory results except for mono, and possibly stereo if stereo impulses were used. Otherwise, for 5.1 and 7.1 and so on, there will be too much correlated reverb in the various speakers, making the sounds too artificial.

Anyone have practical experience with things like the Altiverb? What outcomes did you have relative to using something artificial like a high-end Lexicon reverb?

posted 26 April 2002 07:33            

quote:

---

Originally posted by mathis:

One year ago, when i wanted to program an advanced ambisonics decoder in kyma for soundfield recordings i collected a bunch of aes-preprints and internet articles. .

---

If it's not too much trouble would you mind posting some of this material you collected? I'm very much interested in exploring the possibility of programming a B-Format decoder in Kyma for my Soundfield mic.