Audio Science: Does it explain everything about how something sounds?

[jkeny]

We heard harshness when mastering a recording that had originally been tracked digitally at 16 bits. This was a "triple blind" situation, where we both heard differences from one day to the next when playing the last file the night before as the first file of the next day and it sounded different, even though it was supposed to be the "identical file". It turned out it wasn't, that the dither setting used to produce a 16 bit version had changed due to "cockpit error" on my part. It could be that the harshness came from starting with a recording that had already been subjected to noise shaping as part of the ADC algorithms, but I didn't have visibility into this part of the situation. Levels were those typically used for mixing and mastering, e.g. loud portions at around 85 dB at the listening position. There was no amplifier clipping in this playback volume (due to calibrated tests that guarantee this).

Note that harshness, per se, is not necessarily an indication of sonic distortion in a recording. I've heard it in live choral concerts where I was 10 feet from the singers in the church, so this was a case of my ears distorting. I've occasionally noticed distortion in orchestral concerts in Symphony Hall in Boston, when sitting in row T in the center section of the floor, but only briefly in fortissimo peaks.

I use SoundForge 10c to generate test signals and often use its spectrum plot capability and statistics functions. For sample rate conversions, I use the 64 bit iZotope SRC, presently the version in RX4. This version of the SRC does not add any offset to the sampled waveforms when the linear phase option is selected. Previous versions of this SRC, for example those in RX2 Advanced, have a sub-sample off-set. Versions of this SRC also comes integrated with various software programs, such as SoundForge 10c, but some of these use older versions of the software. (If there is offset in a sample rate conversion then null tests created by mixing out of phase signals will show differences that may not be audible, and variable click artifacts when using defective ABX test software.)

My advice to people is to play with tools and also study the theory of digital signal processing. The two approaches complement each other. I have found plots as appear in technical papers and internet postings can be suspect, because there are too many ways to rig them to show what the creater wished (either mistakenly or intentionally). The only way to recognize these possibilities is hands-on experience backed with some knowledge of theory. None of this came easy for me. I spent quite some time learning the necessary math and even more time playing around with audio editing tools and learning how to hear subtle differences these tools can create. Of course if one just wants to listen to music there is no need for any of this, but if one wants their understanding of the audio hobby or profession to be science based, I don't know of any easy way to get this.

Thanks, Tony, appreciate your input.

 

[PeterA]

I don't follow all of audio science research so I can't give you the full picture. What I can do is give you a strong barometer of it by showing you the titles of papers submitted for the upcoming AES presentation in September. That should give you an excellent idea of what people think is novel enough to come and talk about at this yearly conference.

Since the list is rather long , here is my quick attempt at summarizing most of the buckets:

1. Loudspeakers, loudspeakers, loudspeakers. Whether it is drivers or whole loudspeakers, this area always gets attention as you see from a bunch of papers presented.
2. 3-D sound in all manners from recording, to virtualization (i.e. converting multi-channel to 2-channel), understanding of related science (HRTF, IAD), etc.
3. Sound equalization/room interactions.
4. Speech optimizations/intelligibility.
5. Improving performance of low-bit rate audio codecs (driven by cell/mobile phone application where bandwidth still matters).
6. General signal processing algorithms/improvements (filtering approaches, etc.)
7. Revamping of cinema sound. The current standard, the so called x-curve, needs to be shot in the head and new system of equalization/measurement put in place. Mostly discussed at SMPTE but some also bleeds into AES.
8. Headphones.

Probably a few more areas I am missing but you can look for them in the list:

• Virtual Design of Loudspeakers: From Transducers and Enclosure to System Level

• Low Frequency Behavior of Small Rooms

• Measurement of Low Frequencies in Rooms

Thank you, Amir. That is quite a list and goes to the heart of my inquiry with my OP. I now think that my OP was inelegantly phrased, but it is clear to me now what I suspected when I started this thread, namely, that there is still much to learn.

I isolated three topics from that long list. Virtual Design of Loudspeakers is something that Magico and other speaker manufacturers employ. I don't know anything about it, but it is clear that some manufacturers are doing extensive research in this area and I think, keeping their findings to themselves. I read that Magico, for instance, designs its drivers, crossovers and cabinets primarily through virtual design. They have invested heavily in this capability. They then listen carefully to the results before releasing the product.

The other two areas are fascinating to me for a couple of reasons. First, I have only a small room for my stereo. Low frequencies are an issue in this room, and I have small speakers, with limited response, as a result. Room treatment is problematic because the room is not dedicated to my system but also serves as our living room.

Before seeing these two subjects in the list, I kind of assumed that most everything was already understood about measuring Low Frequencies and about their behavior in small rooms. So at a minimum, here are two subjects in which I am interested, and it is fascinating to learn that researchers are still studying these topics and presenting perhaps new information not widely known or understood by their peers.

This leads me to reiterate what I have written before, namely that there is much room on this forum for both respectful discussions about objective measurements/audio science and shared observations gathered from subjective listening impressions. Of course, there is also room to discuss music, gear, members' systems, and many other topics. I have learned from this thread that there is much more left to learn both by me and by the industry at large.

Very interesting post and list.

 

[rbbert]

...I have learned from this thread that there is much more left to learn both by me and by the industry at large...

Only a non-scientist or a poor engineer would even imagine that the answer to the question that is this thread's title could be "yes".

 

[Orb]

It is known that using the identical dither noise in both channels collapses the sound stage, something that may not be undesirable for a solo instrument recording. This was a problem with early digital work stations where there was insufficient processing power to generate many high quality channels of random noise. This is no longer a problem with modern processors. Mixing in Gaussian noise can mask dither distortion if done at a high level, but TPDF dither can eliminate distortion and noise modulation while adding much less noise.

I don't know why Keith Howard's piano recording sounded harsh. It could be the converters, microphone preamplifiers, microphones, microphone positions, venue acoustics, instrument voicing, or even the performer. (Many pianists are "bangers" and produce a harsh tone.)

The proper test for harshness of dithering is to compare a 24 bit recording with a 16 bit conversion with suitable dither. Adding additional noise (as in Gaussian noise) may serve to mask problems in the original recording and reach incorrect conclusions as to the source of harshness. I was comparing two different dithers and hearing a difference in the harshness, whatever the cause. If one works with digital audio files, or with line level analog signals if one has state of the art converters, one can do these kinds of comparison tests and expect some validity. If one gets involved with speakers and microphones, then one's in an entirely different space of "audio science".

Yeah as I mentioned he revisited this a few years later, unfortunately it is not online but he did say also in the article (see below quote), did you read the whole article or just the part I quoted earlier?

There is no question that noise at this level is audible, or that it can influence sound quality. To appreciate this, you have only to note the development of noise-shaped dither and the strongly held preferences for different noise-shaping algorithms among recording and mastering engineers. Some even eschew all noise-shaped dithers, preferring instead the flat-spectrum alternative. Still, it is a surprise to find that such small changes in noise floor are audible even when flat-spectrum TPDF dither is added to signals that already have a noise floor of higher amplitude.

I don't claim to understand why this should be the case, but my ears tell me it is. I have compared dithered and undithered 16-bit versions of all of the tracks listed in Table 1, and in every case except track 4, using the 24-bit original as the reference, I prefer the undithered version. Certainly the dithered and undithered versions sound different, despite the apparently innocuous nature of the quantization error in the undithered case. When, recently, I tried this out on an audio-industry guest, he expressed the same preference. I also exposed him to another experiment that may offer at least a partial explanation.

What I did for this was to generate five different noise signals—all with the same RMS amplitude but different PDFs—and add them to a piano recording ripped from the European Broadcasting Union's Sound Quality Assessment Material (SQAM) CD. In this track the inherent noise level is about –85dBFS, so the noise was added at an amplitude about 20dB greater in order to swamp it.

But Keith is not the only one to suggest Gaussian like alternatives may be perceived more natural than TPDF.
Cheers
Orb

 

[853guy]

I didn't say unknown unknowns is fiction; I said it is used here as a false wild card to dismiss existing science in the face of data that doesn't agree with perception. I...well Amir would do better...could bring all the data in the world and someone could still answer that while their pet product doesn't test well, not everything can be tested/measured, and it sounds better than the products that do test well because of unknown unknown sonic attributes that cannot yet be measured. OK. In a world where all things are possible, that's one of them. And it could be that their favorite manufacturers are even making stuff that only uses these unknown unknowns for the better, even though they don't know what they are, can't measure them, can't engineer for them, can't identify them and repeat them. Their favorite manufacturers could just have incredible luck.

And we could all be living in the Matrix.

Tim

That’s not what what’s happening. If we can accept there are:

Known knowns: Ohms law, the four laws of thermodynamics, etc
Known unknowns: Uncertainties related to risk, the mechanism causing certain materials to exhibit superconductivity at temperatures higher than 25 kelvin, Galaxy rotation problem, etc
Unknown knowns: Disavowed knowledge or beliefs (torture to sustain public values) in Lacan, Žižek, et, al (Limited to social, moral and ethical research)
Unknown unknowns: “Black Swan” events, etc

then it creates a context in which to define the relationships one has to another within a given body of interest and research. This is the nature of the study of complex systems.

By and large, the quantifiable aspects of our audio systems are the known knowns, and while many of us, myself included, have only cursory knowledge of its processes, we seek greater understanding in light of the known unknowns, namely how our auditory perception perceives music when played back over a complex system (transducer/amplifier/transducer which currently is not capable of perfectly unaltered reproduction due to the laws of thermodynamics) in which its objective analysis was deduced with non-musical signals.

It’s the relationship between the known knowns (Ohms laws, the laws of thermodynamics) and the known unknowns (the relationship between perception, emotion and memory) that is of primary interest to me, because while the objective elements of our audio systems can currently be measured, it still fails to tell me anything about how I will perceive it, as perception of difference is fundamentally different from judgement which involves memory and we are still mapping how emotion affects memory. These latter aspects are the known unknowns.

Ideally, we’d want to build on what we do know, without clinging rigorously to it, by questioning it and critiquing it, in order to explore the relationship between the two because a music delivery device has no purpose for existing unless it is connected to our perception of it. Yes, impedance/THD/IMD must be measured, but music must be perceived. Therefore, to discount perception as a source of continued research is to say the only purpose of an audio system is to produce sine waves and square waves to be measured objectively. (The three independent variable of music - time, pitch, amplitude - can of course be independently measured, but music as an art form is never the constituent parts separated one from another, it’s always the relationship between the three that defines it from sound.)

Bear with me…

I find this parallel with the Black-Scholes-Merton option-pricing formula useful.

Adherents of the Black-Scholes ruled out extreme market events (unknown unknowns) because the option pricing model considered volatility to be relatively constant (a known(ish) unknown). In order for the formula to “work” one needs to input several known knowns (time to maturity, interest rates, strike price, etc) and only one known unknown - the underlying asset’s volatility. The trader must estimate this, and it’s the relationship between the known knowns and the known unknown that the trader seeks to exploit. But there was a problem with this. Black-Scholes assumed large events were outside of all probability. Therefore, traders using the model stuck to the same implied volatility level, with the assumption the option should have just one volatility expectation per maturity date, plotted as a horizontal line (a nice graph!). However, as Black Monday ungraciously reminded everyone, volatility is not and can never be a straight line, it’s much more a smile - an inverse bell curve. As we’ve continued to see, extreme events are very possible, but very unpredictable, and the effect of these events - especially on models who assume linearity in real world conditions - is well documented, as is Long Term Capital Management’s spectacular implosion, of whom Scholes and Merton were on the board of directors. (Despite this, and the fact that they had simply rejigged existing technical and academic research including basically identical formulas, they got to keep their Nobel Memorial Prize in Economic Sciences.)

Trading derivatives is to enter into a system of complexity, which itself has become more complex with the creation of synthetic variations. Scholes and Merton had BSc’s, MA’s, MBA’s, Ph.D.’s and Nobel Prizes. They were fine with the known knowns, and fudged a bit with the known unknowns, but never considered the unknown unknowns. History will not forget them.

How is this related to audio reproduction? Looking only at objectively-sourced measurements based on available data (the known knowns) and making correlations with the still nascent research into the neuro-physiological process of perception, emotion and memory (the known unknowns) is presumptuous to say the least. Personally, I think the relationship between the two has much, much more to be explored. And I’m fully confident robust, ethical scientific research can shed light on this.

Personally, I think discounting the unknown unknowns completely, without acknowledging the possibility they exist outside our current level of knowledge is pure arrogance. Do we know what they are? Not yet. But burying our heads in the sand and focussing only on what we do know isn’t going to get us anywhere, as foolish as it may seem to those content with what they can read in a white paper, even one based on Nobel Prize-winning research.

 

[Tony Lauck]

Yeah as I mentioned he revisited this a few years later, unfortunately it is not online but he did say also in the article (see below quote), did you read the whole article or just the part I quoted earlier?


But Keith is not the only one to suggest Gaussian like alternatives may be perceived more natural than TPDF.
Cheers
Orb

I just read the quoted portion.

The problem with Gaussian noise (specifically white Gaussian noise, samples independently distributed) is one of scale factor. To eliminate distortion (error in average) one can use rectangular dither. This will leave noise modulation (error in second momement). TPDF fixes this. However using a Gaussian probability distribution function won't have these particular values and so will not eliminate distortion and noise modulation, although a close approximation can be had to Gaussian noise by constructing it as the sum of rectangular distributions. Unfortunately, lots of them will have to be summed, resulting in much more noise.

The best you can do with 16 bits is to use subtractive dither (that was invented in the 1970's by Larry Roberts in an MIT Master's Thesis in the context of video). This will eliminate all correlation between any input signal and the quantization distortion. Unfortunately, this can not be used in practical converters. It can be used to compress bit depth in a linear PCM system, and this is done in HDCD. In addition to eliminating all the first and second order correlations that TPDF does, it eliminates third and higher order correlations and also reduces the noise floor by about 6 dB over TPDF. This requires a pseudo random source of synchronized dither. Once one uses subtractive dither the harshness of low bit depth PCM completely disappears, but of course one gets a fair amount of hiss at low bit depths such as 8 bits.

For anyone who is really curious about dither, one can (try to) read the following.
http://robertwannamaker.com/writings/ieee.pdf
http://robertwannamaker.com/writings/rw_phd.pdf

 

[853guy]

Note that harshness, per se, is not necessarily an indication of sonic distortion in a recording. I've heard it in live choral concerts where I was 10 feet from the singers in the church, so this was a case of my ears distorting. I've occasionally noticed distortion in orchestral concerts in Symphony Hall in Boston, when sitting in row T in the center section of the floor, but only briefly in fortissimo peaks.

I've heard much the same thing in choral works too. I remember buying Crystallisatio for chamber orchestra by Estonian composer Erkki-Sven Tüür on ECM. The louder passages exhibited some distortion, which I attributed to mic/pre/converter clipping. But having heard large choral works in a couple of different settings, it seems to be that large choirs and the nature of the human voice's greater modulation at fortissimo, especially when single notes are held for long durations tends to "distort" the sound, kinda like intermodulation distortion, as the combination of each individual voice create sympathetic and unsympathetic harmonics.

That's nothing I can prove, just an anecdote.

 

[jkeny]

orb,
As you probably know by now - this whole area of the misapplication of "thresholds of audibility" has been a suspicion of mine for a while now. I believe both you & Tony have also had these concerns for even longer than I. My focus on noise as one of the most probable areas where the "current audibility thresholds" are not applicable was based on my experience with my own work & reading the work of others. It seems that my recent stumbling upon the ITU-R 468 weighting further reinforces my leanings towards these suspicions. Noise seems to be an overlooked area in audio & taken too literally - it seems that only 1st order effects are the only concern i.e can the noise be heard & not what secondary effect does noise have on the perception of music replay?

There seems to be a great reluctance on the part of objectivists/measureists to engage in this discussion whenever it is raised in posts.

 

[Tony Lauck]

That’s not what what’s happening. If we can accept there are:

Known knowns: Ohms law, the four laws of thermodynamics, etc
Known unknowns: Uncertainties related to risk, the mechanism causing certain materials to exhibit superconductivity at temperatures higher than 25 kelvin, Galaxy rotation problem, etc
Unknown knowns: Disavowed knowledge or beliefs (torture to sustain public values) in Lacan, Žižek, et, al (Limited to social, moral and ethical research)
Unknown unknowns: “Black Swan” events, etc

then it creates a context in which to define the relationships one has to another within a given body of interest and research. This is the nature of the study of complex systems.

By and large, the quantifiable aspects of our audio systems are the known knowns, and while many of us, myself included, have only cursory knowledge of its processes, we seek greater understanding in light of the known unknowns, namely how our auditory perception perceives music when played back over a complex system (transducer/amplifier/transducer which currently is not capable of perfectly unaltered reproduction due to the laws of thermodynamics) in which its objective analysis was deduced with non-musical signals.

It’s the relationship between the known knowns (Ohms laws, the laws of thermodynamics) and the known unknowns (the relationship between perception, emotion and memory) that is of primary interest to me, because while the objective elements of our audio systems can currently be measured, it still fails to tell me anything about how I will perceive it, as perception of difference is fundamentally different from judgement which involves memory and we are still mapping how emotion affects memory. These latter aspects are the known unknowns.

Ideally, we’d want to build on what we do know, without clinging rigorously to it, by questioning it and critiquing it, in order to explore the relationship between the two because a music delivery device has no purpose for existing unless it is connected to our perception of it. Yes, impedance/THD/IMD must be measured, but music must be perceived. Therefore, to discount perception as a source of continued research is to say the only purpose of an audio system is to produce sine waves and square waves to be measured objectively. (The three independent variable of music - time, pitch, amplitude - can of course be independently measured, but music as an art form is never the constituent parts separated one from another, it’s always the relationship between the three that defines it from sound.)

Bear with me…

I find this parallel with the Black-Scholes-Merton option-pricing formula useful.

Adherents of the Black-Scholes ruled out extreme market events (unknown unknowns) because the option pricing model considered volatility to be relatively constant (a known(ish) unknown). In order for the formula to “work” one needs to input several known knowns (time to maturity, interest rates, strike price, etc) and only one known unknown - the underlying asset’s volatility. The trader must estimate this, and it’s the relationship between the known knowns and the known unknown that the trader seeks to exploit. But there was a problem with this. Black-Scholes assumed large events were outside of all probability. Therefore, traders using the model stuck to the same implied volatility level, with the assumption the option should have just one volatility expectation per maturity date, plotted as a horizontal line (a nice graph!). However, as Black Monday ungraciously reminded everyone, volatility is not and can never be a straight line, it’s much more a smile - an inverse bell curve. As we’ve continued to see, extreme events are very possible, but very unpredictable, and the effect of these events - especially on models who assume linearity in real world conditions - is well documented, as is Long Term Capital Management’s spectacular implosion, of whom Scholes and Merton were on the board of directors. (Despite this, and the fact that they had simply rejigged existing technical and academic research including basically identical formulas, they got to keep their Nobel Memorial Prize in Economic Sciences.)

Trading derivatives is to enter into a system of complexity, which itself has become more complex with the creation of synthetic variations. Scholes and Merton had BSc’s, MA’s, MBA’s, Ph.D.’s and Nobel Prizes. They were fine with the known knowns, and fudged a bit with the known unknowns, but never considered the unknown unknowns. History will not forget them.

How is this related to audio reproduction? Looking only at objectively-sourced measurements based on available data (the known knowns) and making correlations with the still nascent research into the neuro-physiological process of perception, emotion and memory (the known unknowns) is presumptuous to say the least. Personally, I think the relationship between the two has much, much more to be explored. And I’m fully confident robust, ethical scientific research can shed light on this.

Personally, I think discounting the unknown unknowns completely, without acknowledging the possibility they exist outside our current level of knowledge is pure arrogance. Do we know what they are? Not yet. But burying our heads in the sand and focussing only on what we do know isn’t going to get us anywhere, as foolish as it may seem to those content with what they can read in a white paper, even one based on Nobel Prize-winning research.

Ohm's law is empirical, based on an approximately linear relationship between voltage and current across certain materials. Things are not so simple as they appeared in the 19th century when the law was originally formulated. So even Ohm's law has its "unknown unknowns". There are various assumptions needed before it will apply, e.g. that the temperature of the material remains constant, something that is not true in practice if current is flowing through a resistor.

I first came across the Black Scholes model in the early 1970's when a friend introduced me to "Beat the Market". This book was published prior to the Black Scholes paper. The principle author, E. O. Thorp, went on to become a memboer of the CBOE and later a professor of mathematics at Princeton. I had some fun trading warrants using his system back in the day. But I didn't for a minute believe that success was guaranteed. Thorp also wrote a book on card counting at blackjack, "Beat the Dealer". This also worked, at least when I tried it, but not having a lot of talent for card memory, I didn't do this very much, just enough to be a small life-time winner at Las Vegas, incidental to a few business trips.

Probability models can prove lots of things and can have enormous errors if the assumptions underneath the model are incorrect. This led me to make an error of a factor of more than 10^20 in the probability of a suspected failure mode of an early Ethernet controller. This failure was "impossible" over the entire life and plausible sales of Ethernet controllers, but the fact was that the failure was observed three days later to happen every three minutes on a test network with three controllers. Lesson Learned...

 

[853guy]

Ohm's law is empirical, based on an approximately linear relationship between voltage and current across certain materials. Things are not so simple as they appeared in the 19th century when the law was originally formulated. So even Ohm's law has its "unknown unknowns". There are various assumptions needed before it will apply, e.g. that the temperature of the material remains constant, something that is not true in practice if current is flowing through a resistor.

Well, see... there you go. Just when I thought I had all my boxes neatly categorised...! Strictly though, wouldn't the variables associated with material temperature be a "known unknown"?

I first came across the Black Scholes model in the early 1970's when a friend introduced me to "Beat the Market". This book was published prior to the Black Scholes paper. The principle author, E. O. Thorp, went on to become a memboer of the CBOE and later a professor of mathematics at Princeton. I had some fun trading warrants using his system back in the day. But I didn't for a minute believe that success was guaranteed. Thorp also wrote a book on card counting at Black Jack, "Beat the Dealer". This also worked, at least when I tried it, but not having a lot of talent for card memory, I didn't do this very much, just enough to be a small life-time winner at Las Vegas, incidental to a few business trips.

Probability models can prove lots of things and can have enormous errors if the assumptions underneath the model are incorrect. This led me to make an error of a factor of more than 10^20 in the probability of a suspected failure mode of an early Ethernet controller. This failure was "impossible" over the entire life and plausible sales of Ethernet controllers, but the fact was that the failure was observed three days later to happen every three minutes on a test network with three controllers. Lesson Learned...

That's a great story, albeit with a slice of humble pie as a side dish. I became fascinated with E.O. Thorp too when I read about his work, but unfortunately I still suck at Black Jack very, very badly. Poker, on the other hand...

 

[rbbert]

...Thorp also wrote a book on card counting at blackjack, "Beat the Dealer". This also worked, at least when I tried it...

There's no question that this works, and for those adherents of free-market economics the ultimate proof is that anyone even suspected of using this or similar systems can count on being barred from entering any casino in the USA.

 

[rbbert]

As far as thresholds of audibility, existing "knowledge" is obviously woefully inadequate. For example, who still believes 1 dB is the JND for loudness, yet it is still held as the standard in Wikipedia and similar references. Funding for perceptual testing in audio is virtually non-existent, which pretty much ensures that little will be done and that what little is done will probably not withstand careful analysis for accuracy or reproducibility. Even in bio-medical research, with lots of funding, studies involving effects on human beings (often measurable, biochemical effects) usually fail to meet fairly simple, standard tests for statistical accuracy and relevance; trying to imagine how it could be as good, much less better, if emotions and perceptions enter the picture boggles the mind.

 

[Orb]

I just read the quoted portion.

The problem with Gaussian noise (specifically white Gaussian noise, samples independently distributed) is one of scale factor. To eliminate distortion (error in average) one can use rectangular dither. This will leave noise modulation (error in second momement). TPDF fixes this. However using a Gaussian probability distribution function won't have these particular values and so will not eliminate distortion and noise modulation, although a close approximation can be had to Gaussian noise by constructing it as the sum of rectangular distributions. Unfortunately, lots of them will have to be summed, resulting in much more noise.

The best you can do with 16 bits is to use subtractive dither (that was invented in the 1970's by Larry Roberts in an MIT Master's Thesis in the context of video). This will eliminate all correlation between any input signal and the quantization distortion. Unfortunately, this can not be used in practical converters. It can be used to compress bit depth in a linear PCM system, and this is done in HDCD. In addition to eliminating all the first and second order correlations that TPDF does, it eliminates third and higher order correlations and also reduces the noise floor by about 6 dB over TPDF. This requires a pseudo random source of synchronized dither. Once one uses subtractive dither the harshness of low bit depth PCM completely disappears, but of course one gets a fair amount of hiss at low bit depths such as 8 bits.

For anyone who is really curious about dither, one can (try to) read the following.
http://robertwannamaker.com/writings/ieee.pdf
http://robertwannamaker.com/writings/rw_phd.pdf

Unfortunately late here and a bit busy.
Yes it is a complex issue and I do recommend TPDF on the forum, but Keith is highlighting an aspect of the various dithers and with some personal listening, if you read the article (I did get the feeling you had not with the first response ) it is pretty clear Keith does feel it is a controversial subject he is raising, and that aspects he raise are not a universal fit to all scenarios.
Worth noting we are talking Gaussian like dither in the context of his articles (shame follow up is not online, but oh man you did not even bother with the one that is at Stereophile )
This is something he has followed up with Tony Faulkner (actually touches a bit of that in this article) and others.
Sorry run out of time to say much more, but I think we would be digressing.
Cheers
Orb

 

[Groucho]

I would be very interested to know the relative levels between acoustic noise generated by power supplies or components on a circuit board in the listening room, compared to the supposedly horrendous din coming from slightly incorrect dither.

http://www.edn.com/design/components-and-packaging/4364020/Reduce-acoustic-noise-from-capacitors

I have certainly heard acoustic noise coming from power transistors as they operated, and I guess the sound might be considered to be equivalent to the dreaded "noise modulation".

If I found out that the acoustic noise levels were comparable to the supposed digital audio errors we're talking about, and yet the Golden Ears hadn't noticed it at all, I might begin to think that this is an imaginary, purely theoretical problem.

 

[Orb]

orb,
As you probably know by now - this whole area of the misapplication of "thresholds of audibility" has been a suspicion of mine for a while now. I believe both you & Tony have also had these concerns for even longer than I. My focus on noise as one of the most probable areas where the "current audibility thresholds" are not applicable was based on my experience with my own work & reading the work of others. It seems that my recent stumbling upon the ITU-R 468 weighting further reinforces my leanings towards these suspicions. Noise seems to be an overlooked area in audio & taken too literally - it seems that only 1st order effects are the only concern i.e can the noise be heard & not what secondary effect does noise have on the perception of music replay?

There seems to be a great reluctance on the part of objectivists/measureists to engage in this discussion whenever it is raised in posts.

Yeah and I feel many of the more vocal objectivists feel there is nothing else to do or relevant; audio sites show this with the response Amir received with his debating at various forums the hirez vs CD ABX pass.
What Stan Curtis did was interesting as well with Discriminant Function Analysis IMO, including his views.
And I think most would agree he is an experienced audio engineer with quite a few successes under his belt.

Cheers
Orb

 

[Groucho]

The final element in all audio amplification is the goal of no distortions of any kind.

Is this true? I agree with you, but I think that a large number of audiophiles indirectly demonstrate that they do not agree - through their espousal of 70 year old technology regardless of its various distorting mechanisms.

Here sine and sine waveforms are still our long suffered and trusted friend.

Sine waves offer a standardised method of quantifying distortion, and allow phase shifts in the system to be measured or (mainly) ignored. But do they show everything? I could imagine an amplifier that exhibits dynamic effects related to momentary power supply dips or thermal effects of components rapidly changing temperature under a sudden change of signal that would not be the same if the amp was under a steady signal (albeit with multi-tone sine waves).

Are sine wave tests of listening rooms meaningful? A delayed reflection will appear as a peak or trough in the frequency response (as measured with sine waves) but this will not be perceived by the listener, who automatically filters out delayed reflections from the direct sound when the source is real music as opposed to sine waves. The impulse response of the room can be derived from a complex sine sweep, however. In this case, the information is all there, but it is the over-simplification of it, presenting it as sine wave magnitudes only, that is the problem.

Do sine wave tests adequately describe the sonic effects of the time domain behaviour of a bass reflex speaker?

 

[jkeny]

Yeah and I feel many of the more vocal objectivists feel there is nothing else to do or relevant; audio sites show this with the response Amir received with his debating at various forums the hirez vs CD ABX pass.
What Stan Curtis did was interesting as well with Discriminant Function Analysis IMO, including his views.
And I think most would agree he is an experienced audio engineer with quite a few successes under his belt.

Cheers
Orb

Yes, orb agree
The DFA stuff is interesting - I will look into it further when I get a chance

 

[Phelonious Ponk]

That’s not what what’s happening. If we can accept there are:

Known knowns: Ohms law, the four laws of thermodynamics, etc
Known unknowns: Uncertainties related to risk, the mechanism causing certain materials to exhibit superconductivity at temperatures higher than 25 kelvin, Galaxy rotation problem, etc
Unknown knowns: Disavowed knowledge or beliefs (torture to sustain public values) in Lacan, Žižek, et, al (Limited to social, moral and ethical research)
Unknown unknowns: “Black Swan” events, etc

then it creates a context in which to define the relationships one has to another within a given body of interest and research. This is the nature of the study of complex systems.

By and large, the quantifiable aspects of our audio systems are the known knowns, and while many of us, myself included, have only cursory knowledge of its processes, we seek greater understanding in light of the known unknowns, namely how our auditory perception perceives music when played back over a complex system (transducer/amplifier/transducer which currently is not capable of perfectly unaltered reproduction due to the laws of thermodynamics) in which its objective analysis was deduced with non-musical signals.

It’s the relationship between the known knowns (Ohms laws, the laws of thermodynamics) and the known unknowns (the relationship between perception, emotion and memory) that is of primary interest to me, because while the objective elements of our audio systems can currently be measured, it still fails to tell me anything about how I will perceive it, as perception of difference is fundamentally different from judgement which involves memory and we are still mapping how emotion affects memory. These latter aspects are the known unknowns.

Ideally, we’d want to build on what we do know, without clinging rigorously to it, by questioning it and critiquing it, in order to explore the relationship between the two because a music delivery device has no purpose for existing unless it is connected to our perception of it. Yes, impedance/THD/IMD must be measured, but music must be perceived. Therefore, to discount perception as a source of continued research is to say the only purpose of an audio system is to produce sine waves and square waves to be measured objectively. (The three independent variable of music - time, pitch, amplitude - can of course be independently measured, but music as an art form is never the constituent parts separated one from another, it’s always the relationship between the three that defines it from sound.)

Bear with me…

I find this parallel with the Black-Scholes-Merton option-pricing formula useful.

Adherents of the Black-Scholes ruled out extreme market events (unknown unknowns) because the option pricing model considered volatility to be relatively constant (a known(ish) unknown). In order for the formula to “work” one needs to input several known knowns (time to maturity, interest rates, strike price, etc) and only one known unknown - the underlying asset’s volatility. The trader must estimate this, and it’s the relationship between the known knowns and the known unknown that the trader seeks to exploit. But there was a problem with this. Black-Scholes assumed large events were outside of all probability. Therefore, traders using the model stuck to the same implied volatility level, with the assumption the option should have just one volatility expectation per maturity date, plotted as a horizontal line (a nice graph!). However, as Black Monday ungraciously reminded everyone, volatility is not and can never be a straight line, it’s much more a smile - an inverse bell curve. As we’ve continued to see, extreme events are very possible, but very unpredictable, and the effect of these events - especially on models who assume linearity in real world conditions - is well documented, as is Long Term Capital Management’s spectacular implosion, of whom Scholes and Merton were on the board of directors. (Despite this, and the fact that they had simply rejigged existing technical and academic research including basically identical formulas, they got to keep their Nobel Memorial Prize in Economic Sciences.)

Trading derivatives is to enter into a system of complexity, which itself has become more complex with the creation of synthetic variations. Scholes and Merton had BSc’s, MA’s, MBA’s, Ph.D.’s and Nobel Prizes. They were fine with the known knowns, and fudged a bit with the known unknowns, but never considered the unknown unknowns. History will not forget them.

How is this related to audio reproduction? Looking only at objectively-sourced measurements based on available data (the known knowns) and making correlations with the still nascent research into the neuro-physiological process of perception, emotion and memory (the known unknowns) is presumptuous to say the least. Personally, I think the relationship between the two has much, much more to be explored. And I’m fully confident robust, ethical scientific research can shed light on this.

Personally, I think discounting the unknown unknowns completely, without acknowledging the possibility they exist outside our current level of knowledge is pure arrogance. Do we know what they are? Not yet. But burying our heads in the sand and focussing only on what we do know isn’t going to get us anywhere, as foolish as it may seem to those content with what they can read in a white paper, even one based on Nobel Prize-winning research.

Actually, that is what's happening here. Most using the "unknown stuff" argument here don't know and don't understand what you posted here, they just use the "science doesn't explain everything" argument to dismiss or marginalize things that science does explain, when it isn't supporting what they believe. Follow the posts. This argument doesn't come up because people are simply curious about the possibilities; it comes up when science has been presented, and has cast doubt on pet theories and darling products. And by contrast, I don't think anyone here has said we already know everything there is to know about audio reproduction or human hearing and perception, in spite of how often that has been repeated. But we're pretty sure about some things. Some things have been tested, reviewed and established for decades, by people working in much more critical fields than audio reproduction, yet they are consistently denied in the audiophile world. Is there a possibility that something, someday will be discovered that will change what we currently understand? Of course. But it won't, for example, make speakers with a narrow sweet spot sound good off axis. And it won't make them sound good in the sweet spot when those first reflections reach the listener. It won't change the results of blind listening tests that rated those speakers below much less expensive models with smooth off-axis response. When that happens, someone questions the validity of the methodology, the system's ability to drive the speakers, the objectivity of the researchers...with impunity, but without substance. And so it goes.

Enough of this. It is an endless loop and not only will no one change their minds, they will characterize the minds and positions of those who disagree with them to fit the arguments they want to make, and when that fails, they will imagine unknown unknowns that support their preferences and positions where the existing science does not. The only way to end it is to stop doing research that slays sacred cows. Or stop discussing it. On a discussion forum.

Tim

 

[Phelonious Ponk]

Noise seems to be an overlooked area in audio & taken too literally - it seems that only 1st order effects are the only concern i.e can the noise be heard & not what secondary effect does noise have on the perception of music replay?

There seems to be a great reluctance on the part of objectivists/measureists to engage in this discussion whenever it is raised in posts.

Start a thread, Jon. I'd be interested.

Tim