or Connect
Head-Fi.org › Forums › Equipment Forums › Sound Science › measurements are models
New Posts  All Forums:Forum Nav:

measurements are models - Page 3

post #31 of 93

I suspect someone is bored and decided to start a troll post. if measurements were not usefull and models not good enough, try to picture in your head the world without those 2 "inaccurate" and "useless" tools...

 

I guess books are useless they don't depict reality perfectly, it's a bad model.

this is getting dumb.

 

 

 

 

 

 

 

 

 

 

 

guys you need to show me the way of Zen. you seem to never lose your cool

 

post #32 of 93

The important part to take away is that nothing is 100% perfect, but that does not make it useless.

 

Models you make in your brain are far from perfect. What you perceive does not match 100% objective reality. The big problem with subjectivity is reproducibility.

 

Example: If you see the cable you're listening with then you know which cable it is. You might hear a difference to your other cable. Why? Your brain creates a model of exactly that cable since you know which cable it is, therefore altering your perception accordingly.

Now try to prove that there indeed is a difference: hide the cables. Now all your brain can make a model of is the sound and only the sound. If you cannot hear the difference anymore then your brain created a truly faulty model of the cable in the sighted case. If you can hear a difference then maybe the model was pretty close to reality, BUT measurements will show those differences and they will be trivial to point out.


Edited by xnor - 12/22/13 at 3:31am
post #33 of 93
Thread Starter 
Quote:
Originally Posted by dvw View Post
 

You are confusing the design parameter with hearing model. For cost efficiency reason, an audio design goal is to be able to address the hearing of 99% of the people. A physiological model can only be a statistical model because of aging and anatomical deviation. A large portion of the population can't hear over 16KHz. So we can't even predict how useful these bandwidth is to all people. A prediction of the brain's response is not only difficult, but it's also unnecessary. Each person might response to an stimulant differently. But if the stimulant is an identical copy, an identical response should result. This is kind of like ABX. You measured the differential and not the absolute value.

 

Models can be useful in making certain prediction. But for design purpose, it can also use to uncover flaws that can be costly or deadly.


I agree with what you write here; I think we are using the word "model" a little differently. I'm deliberately using it loosely for a very important reason: to show how pervasive models and PREDICTION by models is. You are going into much more detail about the type of prediction, distinguishing between statistical models etc.

 

The cochlea's response can be partially modeled by describing its bandwidth, can it not?

post #34 of 93
Thread Starter 
Quote:
Originally Posted by xnor View Post
 

When speaking of models and prediction I think of dynamic systems that change and models that try to predict those future changes. Audio measurements are usually not of that nature.

 

 

 

 

Ah, but audio system are dynamic and we do consider their future behavior. When we measure an audio device, we have a way to try to predict its future response to a given input signal. Our measurement may be more or less useful depending on what this future signal is. The signal running through an audio device is always changing while that device is playing music. The hardware doesn't change, but activity of the hardware changes, and that's what we are trying to predict.

 

Quote:
 What it it useful for? Well, you can see the roll-of at the extremes. With SS amps you can see signs of instability. Sometimes you can see a bass boost and/or treble boost that cannot be disabled, which is rare but exists. Or you can see what a switchable bass boost does exactly. With DACs you can often tell whether it is oversampling or not. You can see the filter (ideally you'd also have the phase response) that is used. With headphones you can see, depending on the level of averaging, balance between lows/mids/highs, resonance problems, nasty treble peaks etc.
 
 
 

 

You're taking a narrow view here, in answering the question "What is it useful for?" You say you can see signs of instability; I could then ask "What use is seeing signs of instability?" "What use is knowing whether a DAC is oversampling?" Think more generally: what is the purpose of an audio system?

 

Quote:
What does it predict? The measurement itself predicts nothing. But I can predict that a basshead will probably enjoy the sound signature of a headphone with boosted bass more than one with flat or weak bass that rolls off early, etc. 

 

If a measurement predicts nothing, then it's a useless measurement.

 

But your comment about the basshead is the first time you've referred to how a system is perceived, which is the real key and real purpose of an audio system.

post #35 of 93
Quote:

For what purpose should the captured signal and its reproduction be as similar as possible? In other words, what is the usefulness of achieving this? You don't seem to be thinking about that at all, despite the fact that the banner of the Sound Science forum refers to models and usefulness.

 

"Recording equipment, playback, etc." are not models. Those are devices. It's not clear that you even know what a model is.

 

 

I couldn't care less for the banner. In my point of view its an incredibly narrow definition of science, written by someone whose own methods and assertions have been criticized in the scientific community. Maybe you should've rephrased the topic of your thread: "Do you agree with the banner of the Sound Science forum?".

 

The usefulness of achieving accurate reproduction? There are other more important uses for signals, the audio scene is a by product of which. For example the signals of the keyboard you typed this message on, and the signal that transmitted it to the forum's server.

 

I know what I listed were devices constituting a model, I was talking about a model that captures and reproduces signals. Its the same model everywhere, except the capturing and reproducing components differ.

 

And now that you've deemed I'm incapable of understanding a model, I think I'll let those who stand by the banner take it from here. May the best model win.


Edited by proton007 - 12/22/13 at 4:24am
post #36 of 93
Quote:
Originally Posted by raddle View Post
 

 

 

Ah, but audio system are dynamic and we do consider their future behavior. When we measure an audio device, we have a way to try to predict its future response to a given input signal. Our measurement may be more or less useful depending on what this future signal is. The signal running through an audio device is always changing while that device is playing music. The hardware doesn't change, but activity of the hardware changes, and that's what we are trying to predict.

I see things differently. By the time we've measured let's say FR it's already "outdated" and if you insist on it being a "model" then such a measurement really would be just a descriptive model of what happened in the past. Since, for all intents and purposes and to make things simpler, audio devices do not change the past observation is still valid now and in the future.

 

If you compare this to gravity, then measurements would be the speed of an object or distance it traveled since it was dropped at different points in time. That's no predictive model. The model would be formulas such as h(t) = h0 - 1/2*g*t^2.

 

A model that can really predict the output for any input signal would be, for example, a guitar amp model - a complex mathematical model that takes any input signal and transforms it into a distorted output signal. The better the model, the closer the output will be to the real thing.

 

 

Quote:
You're taking a narrow view here, in answering the question "What is it useful for?" You say you can see signs of instability; I could then ask "What use is seeing signs of instability?" "What use is knowing whether a DAC is oversampling?" Think more generally: what is the purpose of an audio system?

I just gave you practical examples, no superficial babble. If the amp is unstable then you should return it. Now you could continue asking "what is the use of returning an unstable amp?" and I'd answer "because I don't want a broken piece of crap that might harm my speakers/headphones" and so on.

 

The purpose (as in the target) is faithful reconstruction. Roll-off, instability, bad filters, non-oversampling, funky FR etc. are all signs of failing to meet objective high fidelity reproduction criteria.

 

 

Quote:

If a measurement predicts nothing, then it's a useless measurement.

Disagree. Measuring my weight doesn't predict anything. Measuring my height doesn't predict anything. But combining both can tell me roughly if I'm healthy, too thin or too fat.

 

 

Quote:
But your comment about the basshead is the first time you've referred to how a system is perceived, which is the real key and real purpose of an audio system.

Obviously an audio rig reproduces music for us to perceive it. But that's a bit like saying an atomic clock is made for us to see the time like a wristwatch, duh, but the real purpose is extremely high timing accuracy.


Edited by xnor - 12/22/13 at 9:28am
post #37 of 93
Quote:
Originally Posted by raddle View Post
 

 

If a measurement predicts nothing, then it's a useless measurement.

 

 

 I don't understand why measurement needs to predict anything. Measurement is a verification tool. Your height and weight does not predict anything, but are they useless. Do you ever use the clock and calender? They are a measure of time and they predict nothing. The temperature outside predict nothing either. But we do rely on them as an indicator and use it to control our heating system.

 

I don't know where you're going with this, From your first post, somehow it appears you're trying to discount physical measurement and replace it with "human perception" measurement. And the discussion got way off topic.

post #38 of 93
Quote:
Originally Posted by raddle View Post
 

 

 

Ah, but audio system are dynamic and we do consider their future behavior. When we measure an audio device, we have a way to try to predict its future response to a given input signal. Our measurement may be more or less useful depending on what this future signal is. The signal running through an audio device is always changing while that device is playing music. The hardware doesn't change, but activity of the hardware changes, and that's what we are trying to predict.

 

 

You're taking a narrow view here, in answering the question "What is it useful for?" You say you can see signs of instability; I could then ask "What use is seeing signs of instability?" "What use is knowing whether a DAC is oversampling?" Think more generally: what is the purpose of an audio system?

 

 

If a measurement predicts nothing, then it's a useless measurement.

 

But your comment about the basshead is the first time you've referred to how a system is perceived, which is the real key and real purpose of an audio system.

This stems from a fundamental misunderstanding of science, models, and measurements.

 

A measurement is an observation of reality.

 

A listening test (even an unblinded one) is no less a 'measurement' as using a calibrated microphone to measure the magnitude response at 1 kHz is.

 

The microphone measurement is more objective than the listening test because if you repeat the same test at a different time, with a different person, or in a different place, you'll get a similar result every time with the microphone while the unblinded listening test will vary widely depending on many other factors (e.g. who the listener is, what music is playing, the listener's mood).

This is the definition of 'objective' (versus subjective).

 

Neither of these are models. A model is a theoretical construct, which takes an input and produces an output which fits previous measurements and (hopefully) makes correct predictions.

 

For example, a model from a listening test could be "shiny earphones sound better".

These models may be accurate or inaccurate.

post #39 of 93
Thread Starter 
Quote:
Originally Posted by higbvuyb View Post
 

This stems from a fundamental misunderstanding of science, models, and measurements.

 

A measurement is an observation of reality.

 

 

 

 

 

Yes, it's an observation of reality, but made with imperfect devices, containing noise, etc. Anything from measuring a voltage to measuring a frequency response involves taking some data and then fitting it to a model of how the test equipment behaves. If you disagree then describe to me a truly direct observation. I don't know of any.

 

Quote:
 ~~A listening test (even an unblinded one) is no less a 'measurement' as using a calibrated microphone to measure the magnitude response at 1 kHz is. The microphone measurement is more objective than the listening test because if you repeat the same test at a different time, with a different person, or in a different place, you'll get a similar result every time with the microphone while the unblinded listening test will vary widely depending on many other factors (e.g. who the listener is, what music is playing, the listener's mood). This is the definition of 'objective' (versus subjective).

 

Actually I agree with the notion that a listening test is a measurement. Very good generalization you have there.

 

Quote:
 ~~Neither of these are models. A model is a theoretical construct, which takes an input and produces an output which fits previous measurements and (hopefully) makes correct predictions.

 

Here's why a frequency response is a model. You put a signal into the device and measure its output. You then use some mathematics to computer the FR. In using these mathematics, you are assuming the device is a linear. You are fitting the data to a generalized model, and then coming up with a more specific model for that device.

 

FR also predicts future observations. It predicts the result of putting brand-new, never-used signals into the device. Whether these are accurate predictions depends on how linear the device is.

 

Incidentally, FR is a model of subjective hearing also, a rough one. The ear can hear peaks and valleys in the FR. It's sensitive to the overall width.

post #40 of 93
Quote:
Originally Posted by raddle View Post
 

 

 

Yes, it's an observation of reality, but made with imperfect devices, containing noise, etc. Anything from measuring a voltage to measuring a frequency response involves taking some data and then fitting it to a model of how the test equipment behaves. If you disagree then describe to me a truly direct observation. I don't know of any.

That isn't really a 'model'. In that example, the measurement is still just the raw data set which is a pretty direct measure of what the number on the screen said at the time you observed it. Yes, you could argue that with some definition of 'direct', nothing is direct, but then a category that doesn't apply to anything isn't actually useful.

 

The 'model' you are referring to is what you use when you try to interpret this data with that model.

 

On another note equipment used within the bounds it was designed for for a reasonable degree of accuracy, the model in use is very rigorously tested, so it isn't really an issue. Unfortunately our 'hearing' isn't so reliable.

 

 

FR is nothing like a model of hearing. You can apply a model of hearing to the FR, that's all.

post #41 of 93
Thread Starter 
Quote:
Originally Posted by dvw View Post
 

 I don't understand why measurement needs to predict anything. Measurement is a verification tool. Your height and weight does not predict anything, but are they useless. Do you ever use the clock and calender? They are a measure of time and they predict nothing. The temperature outside predict nothing either. But we do rely on them as an indicator and use it to control our heating system.

 

I don't know where you're going with this, From your first post, somehow it appears you're trying to discount physical measurement and replace it with "human perception" measurement. And the discussion got way off topic.

 

People tend to lose sight of the fact that measurements are models when the situation is simpler or less consequential.

 

If you report to me that your weight is 160 and your height is 5'10", you are using a model. That's because your height and weight change over time. (You are taller in the morning, for instance.) You are modeling them as if they were constants. And sometimes that's good enough.

 

These numbers are also predictions of your height and weight at a future time. But they will never be perfectly accurate. They are useful however, but how useful they are depends on (1) how you computed them, and (2) what you intend to use them for. If you intend to use your weight to determine if it's safe to use a parachute with a maximum weight load, well you may want to investigate how much tolerance is built into the parachute's rated maximum and that kind of thing. If you are using your weight to determine your BMI, well BMI is not a precise predictor of anything (not to the decimal point) so there is much less consequence to an inaccurate weight measurement.

 

But the essence of modeling is here. The model uses assumptions, involves fitting data, makes predictions, and has a usefulness which depends on context.

 

It just looks a little funny because people are used to talking about their height and weight as fixed, known, easily-measureable quantities. It's just a habit of perspective.

post #42 of 93
Quote:
Originally Posted by raddle View Post
 

 

People tend to lose sight of the fact that measurements are models when the situation is simpler or less consequential.

 

If you report to me that your weight is 160 and your height is 5'10", you are using a model. That's because your height and weight change over time. (You are taller in the morning, for instance.) You are modeling them as if they were constants. And sometimes that's good enough.

 

These numbers are also predictions of your height and weight at a future time. But they will never be perfectly accurate. They are useful however, but how useful they are depends on (1) how you computed them, and (2) what you intend to use them for. If you intend to use your weight to determine if it's safe to use a parachute with a maximum weight load, well you may want to investigate how much tolerance is built into the parachute's rated maximum and that kind of thing. If you are using your weight to determine your BMI, well BMI is not a precise predictor of anything (not to the decimal point) so there is much less consequence to an inaccurate weight measurement.

 

But the essence of modeling is here. The model uses assumptions, involves fitting data, makes predictions, and has a usefulness which depends on context.

 

It just looks a little funny because people are used to talking about their height and weight as fixed, known, easily-measureable quantities. It's just a habit of perspective.

The measurement is that you using this particular measurement procedure at this time, you get a result of 5'10". This is a constant.

 

Assuming that your height stays constant over some period of time would be a model (a poor one).

post #43 of 93
Thread Starter 
Quote:
Originally Posted by higbvuyb View Post
 

That isn't really a 'model'. In that example, the measurement is still just the raw data set which is a pretty direct measure of what the number on the screen said at the time you observed it. Yes, you could argue that with some definition of 'direct', nothing is direct, but then a category that doesn't apply to anything isn't actually useful.

 

The 'model' you are referring to is what you use when you try to interpret this data with that model.

 

 

 

 


I think your view and understanding of models is too narrow. Note that models come in different forms, and some are more specific and parameterized, while some contain unknown variables.

 

For example, you can do the math in a linear systems textbook, and you will find that you can make a lot of predictions about how linear systems behave that are generalized to all linear systems. You don't need to know the FR specifically. That's one kind of model.

 

Measuring FR of a specific device involves taking the math and using it to fit observations, and compute the FR. What you end up with is a more specific model.

 

 

 

Quote:
 ~~On another note equipment used within the bounds it was designed for for a reasonable degree of accuracy, the model in use is very rigorously tested, so it isn't really an issue. Unfortunately our 'hearing' isn't so reliable.

 

First of all, you aren't arguing that measurements aren't models, you are just arguing that they are very good models. Sometimes they are. As far as models of audio equipment, they are only useful insofar as they have been "rigorously" correlated with hearing.

 

Quote:
~~FR is nothing like a model of hearing. You can apply a model of hearing to the FR, that's all. 

 

Of course it's a model of hearing. An FR with a big peak in the bass is roughly correlated with the listening experience we report as "loud bass." It's a rough model, but this is a very important fact, because as I said, models in audio are only useful insofar as they have been correlated with the listener's experience.

post #44 of 93
Thread Starter 
Quote:
Originally Posted by higbvuyb View Post
 

The measurement is that you using this particular measurement procedure at this time, you get a result of 5'10". This is a constant.

 

Assuming that your height stays constant over some period of time would be a model (a poor one).

Assuming it's constant might be a perfectly useful model. It would be for computing BMI, for instance. Don't assume it's a bad model.

 

It depends on what use you want to put the measurement to.

 

Very simple idea, often overlooked.

 

The number 5'10" is not separate from the context within which it was computed, and knowing that context helps to make predictions.

 

That number was computed with a device that had error and noise, for instance. If I know how much error and noise, it helps future predictions.

 

If I know whether you measured that in the morning or evening, it helps future predictions.

 

It's really simple, folks. All measurements are models.

 

You can argue that it doesn't matter much, but in audio it does matter a whole heck of a lot.

post #45 of 93
Thread Starter 
Quote:
Originally Posted by xnor View Post
 

 

 

A model that can really predict the output for any input signal would be, for example, a guitar amp model - a complex mathematical model that takes any input signal and transforms it into a distorted output signal. The better the model, the closer the output will be to the real thing.

 

 

 

 

You are confusing the question of whether a model predicts something with the question of whether it predicts it accurately.

 

FR is a way of predicting the output a device given a brand-new input signal, never previously measured.

 

Since this assumes the device is linear it may be more or less accurate.

 

A model that includes nonlineariies would be more accurate.

 

It all depends on what you want to use it for. If you want to determine if an amp sounds bassy, then a good start is to measure FR.

New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Sound Science
Head-Fi.org › Forums › Equipment Forums › Sound Science › measurements are models