"Log sine sweeps rather than linear sine sweeps were employed to allow verification that non-linear distortion components were virtually absent."
And with that, this study is bullshit.
Human beings don't listen to linear sine sweeps. We listen to music. Recorded music has 8+ octaves of frequency range (the bottom octave plus a little extra is almost always rolled off in real-world recordings, to ease stress on downstream components that can't reproduce such low frequencies anyway), and 20-50db of useable dynamic range.
Sine wave measurements of audio gear ignore impulse response, intermodulation distortion, phase shift, and a host of other real-world physical device responses to real-world musical signals. Scientific, reductionist thinking is inadequate to get an accurate picture of the factors that matter to human listeners.
Frequency response and total harmonic distortion aren't measured in these cases because they're useful or relevant. They're measured because they're easy to measure. It's like looking in the wrong place, because the light is better there. And the results? It's like measuring a car's performance by how well it can drive in a straight line at 60mph. Acceleration, braking, and turning are too hard to measure, so we ignore them...
I'm a musician and record producer. I've engineered and produced numerous albums, and rely on multiple different types of headphones for different purposes. The article's claim that one headphone can be easily morphed into another through mere equalization is, frankly, bullshit. The two headphones I rely on the most (Beyerdynamic DT880 and AKG K240) sound wildly different. Neither is "accurate". Neither are the Tannoy System 12 DMT midfield studio monitors I use for mixing, or the stock Subaru car speakers I use for reference to check the mixes from the Tannoys.
Audio reproduction is incredibly complex and difficult stuff. Trying to isolate one factor and saying "That explains everything!" is bad thinking.
> Human beings don't listen to linear sine sweeps. We listen to music. ...
This is irrelevant. They're measuring frequency response, not trying to map the entire world of psychoacoustics. Log sin wave sweeps are a perfectly adequate way of measuring a transducer.
> Sine wave measurements of audio gear ignore impulse response, intermodulation distortion, phase shift, and a host of other real-world physical device responses to real-world musical signals.
No, when you do a sin sweep and measure the impedance you can recover all that. Pick up a basic textbook on loudspeaker testing.
> Scientific, reductionist thinking is inadequate to get an accurate picture of the factors that matter to human listeners.
Learning from it will require you to shed the chip on your shoulder, but it will put you miles ahead of your peers in understanding what's actually going on when we listen to music vs all the commonly repeated mythology.
I second this recommendation. Chapter Two of Dr. Toole's book is titled "Preserving the Art" and directly addresses the concerns you're raising. He uses a great quote at the end of this chapter:
"Not everything that can be counted counts, and not everything that counts can be counted".
- Einstein
Dr. Toole then spends 500 pages addressing how you can correlate objective and subjective evaluation of loudspeakers. It's a great read.
> No, when you do a sin sweep and measure the impedance you can recover all that. Pick up a basic textbook on loudspeaker testing.
Can you really quantify nonlinearities from a sin sweep of any sort? Not that I'd expect a decent set of headphones to have material nonlinearities unless seriously overdriven.
> Can you really quantify nonlinearities from a sin sweep of any sort? Not that I'd expect a decent set of headphones to have material nonlinearities unless seriously overdriven.
You can using a properly structured log sine sweep but not with a linear sine sweep. This is why the paper mentions "Log sine sweeps rather than linear sine sweeps were employed to allow verification that non-linear distortion components were virtually absent."
Not that I'd expect a decent set of headphones to have material nonlinearities unless seriously overdriven.
"Material" is just a way of dismissing the current imperfections that we can't economically do anything about. Headphones and speakers are certainly good enough to let us enjoy music. But listening to recordings on them is like watching cinema. It's certainly enjoyable, but no one thinks it's a duplicate of the real thing. They're just the performance hardware of particular media.
> "Material" is just a way of dismissing the current imperfections that we can't economically do anything about.
I said material nonlinearities, not material imperfections. I've never researched this for real, but I would imagine that your average imperfect set of headphones is mainly imperfect in that it has poor (or inconsistent) frequency response, poor transient response, and/or poor noise characteristics. But maybe headphones do commonly have significant nonlinearities -- I don't really know.
I said material nonlinearities, not material imperfections.
Precisely what I said.
But maybe headphones do commonly have significant nonlinearities
Dynamic drivers have these. Most cases and housings introduce these. Amplifiers have these. Until recently we could only determine if people can consciously hear things or not. Determining if something affects your perception is something else entirely. Also, with something you wear like a headphone, it's more expensive to do blind comparisons.
Basically, the system we're dealing with (including the human hearing apparatus) is pretty complex. While we've made significant inroads into the science, we still don't understand everything that goes on when someone listens to reproduced music. I think this is about to change, however. There are some things coming down the pike, where we are going to have unprecedented control of variables we didn't even think to address.
Here's one of the first baby steps: A headphone that does a frequency sweep inside the cups to compensate for the acoustic filtering of your particular ear!
Sine sweeps are a good method to measure frequency response. That's what they were trying to do after all: measure frequency response. That was the stated goal of the test: find a correlation between frequency response and another parameter. They were playing a signal into a dummy, capturing the sound and measuring it; they were not trying to determine how pleased the dummy would be with the beautiful music.
However, I share your overall assessment that present-day transducers are far from perfect accuracy.
Sure, but the non-linear effects he's listing are probably not an issue for headphones. First, they're generally crossover-less and usually single-driver designs and they have very low moving mass which eliminates a host of issues that loudspeakers have. All those things are still there, and probably quantifiable, but I'd imagine that they're all low enough in headphones to be inaudible or 'not a problem.' What absolutely is a problem is manufacturer fuckery with frequency response ala 'megabass' and the like which is what really moves headphone sales. So it makes sense to me that that is what this study focuses on
> What absolutely is a problem is manufacturer fuckery with frequency response ala 'megabass'
Yep.
Now, that happens in any industry where consumers don't have an easy way to quantify what they're buying. TV screens in showrooms all have color saturation turned way up. Etc.
It should be said, however, that "megabass" or a V-shaped frequency response tends to be the case with intensely advertised, mass produced brands. The kind of stuff you find at Best Buy. It is much less of an issue, or not an issue at all, with brands aimed at people who pick transducers specifically for accurate sound.
E.g. I would expect to encounter "manufacturer fuckery" with Bose or Beats phones, and the like. I would not expect to encounter that with Sennheiser (at least their top models), AKG, Beyerdynamics, Audeze, Hifiman, Stax, Focal, Mr. Speakers, Oppo, etc. Okay, maybe a bit with Beyer. :)
That's not to say that some of these brands are not messing with the response at all. But when they do it, they do it in very subtle ways. The frequency response might be super-flat (or, rather, following the Harman curve which is appropriate for transducers placed on your ears). But they might "brighten" the sound a little, very subtly.
And then when another manufacturer comes along and says "screw those tricks, let's aim for perfect honesty instead", they sound a bit "veiled" in comparison. I'm talking about this in much greater detail here:
Anyway, we are still pretty far from perfect accuracy when it comes to transducers. All other parts of the audio chain have been figured out long ago. But transforming sound into electricity and vice-versa is still a hard problem.
Agreed, but I have returned a pair of sennheiser phones because they butchered the sound too much ... think it was one of the first in the momentum line if I recall
Well, when it comes to Sennheiser, there's the HD line and then there's everything else.
Even then, I think they really, really care about the HD600 / HD650 / HD800 succession of flagships. They might care a little bit about classic workhorses such as HD280 and the like. The rest I feel they are treating a bit more like crowdpleasers and moneymakers. I could be wrong.
It bothers me when they call them "monitors" and then go ahead and fiddle with the frequency response anyway. Look at the difference between the HD 280 pro and HD 380 pro, both labelled as monitoring headphones. The 380 sounds like a muddy mess in comparison, with everything below around 300hz turned up way too high. I think it's fair to ask: if you're going to make MegaBass headphones, mark them as such. Don't advertise flat response and then hand us this.
It absolutely looks to me like you're just trying to keep this field unscientific. Misrepresenting the article, then throwing out a bunch of jargon without explanation, then coming to the conclusion that people here shouldn't even try to understand.
The study doesn't say "that explains everything!". It measures one component in isolation. That's what good science does. The existence of other factors in headphone performance is irrelevant.
Besides, would you really not consider it news if 0-60 time in cars didn't correlate with price? Yes, a luxury car can justify its price in many other ways, but, on average, there should still be a correlation.
I wasn't talking measuring 0-60 time in cars. I was talking about measuring steady 60mph. No one does that, because it's useless (although it would be easy to measure). 0-60 time would correlate somewhat to impulse response in audio, which is a useful measurement, and much more rarely seen.
I'm reminded of an experimental example a professor showed us in my college audio recording class. He hooked a square wave generator up to both analog tape and digital recorders, and recorded the square wave. Then he played back both sources on an oscilloscope. The tape had a sharp lead spike followed by a shrinkage on the tail of the square, from compression. The digital recording had a rippled top, from frequency-related distortion. Both were inaccurate, but they were inaccurate in very different ways.
You can't accurately reproduce a square wave by using a Fourier series even mathematically in the infinitely wide frequency response limit, ignoring all physics:
Why are we talking about Fourier series? Uncompressed digital audio doesn't need them, it just records the raw voltage level samples. You still can't produce a perfect square wave, but there's no reason you shouldn't be able to reproduce whatever wave was provided as input (which by definition was produced by something), other than hardware limitations of the playback system.
You are quite wrong about that statement. As the other comment mentioned, both analog and digital systems are bandwidth limited and can not reproduce infinite slopes like in a square wave. Fourier expansions are one of the easier ways to study bandwidth, even if there are other formalisms.
But that's the analog system which is band limited, not the digital! The problem is poor analog components after the digital decode phase in this case. A time domain digital representation absolutely can represent a perfect square wave. (There are other waves it can't represent.) That's completely different from a digital encoding that causes ringing in the square wave.
A DAQ that reconstructs a perfect square wave (or a perfect stair-step function) is employing a "Zero Order Hold", and it would make a god-awful audio DAC.
With a 0OH you'd gain the ability to potentially reconstruct a infinite bandwidth signal (the squarewave), which is irrelevant in the audio case: With a 10kHz squarewave, the next harmonic would be at 30kHz which is already waaaay out of the human range of hearing.
But limiting the bandwidth of a DAC (or ADC) is the fundamental property which causes it to be able to perfectly reproduce the frequencies within its bandwidth (~up to 20kHz in HiFi Audio), so that's desirable. And the "ripples" on top of a squarewave are just the manifestation of this property: If you cut away the frequencies above Nyquist, you get a rippled squarewave. And conversely if you compute the difference between the perfect squarewave and the "rippled" squarewave you get out of an Audio DAC this only contains (non-audible) energy outside of the bandwith of the DAC!
Sure. But you're still talking about the analog portion of the circuit. If the analog output of a digital recording - supposedly the input was a near-perfect square wave - is different from the analog output of an analog recording, and neither looks like the input signal, then it is a) certainly possible to make an analog output stage that produces a more precise output that better matches the input, and b) possible to make an output that better matches the analog output. Remember, the input was supposedly a perfect square wave, and contained inaudible components. The recording/playback component had nothing to do with the fact that you can't hear the entire spectrum.
All the limitations are in the analog phase. As you point out, it depends on the design tradeoffs in the DAC, amplifiers, etc, and that's an important lesson to learn in the class that was being taught. Nevertheless, the point I was replying to is the claim that the digital representation could not represent a square wave. That's certainly not true, and no Fourier transforms are necessary to demonstrate it. A PCM recording is just a series of impulses, not a series of sine waves.
Inaccuracies visible on an oscilloscope might not be audible to humans. Try applying a 10Hz high-pass filter to some music and comparing the waveforms. There's obvious visible changes but it sounds identical.
True. And it certainly doesn't explain "better" or "worse". But it explains why music recorded on tape sounds different from music recorded digitally. And it puts a torch to the claims that digital is somehow "accurate", and all we like about analog is just evil bad naughty distortion.
Square waves are impossible to reproduce because they have infinite numbers of frequencies, but you don't listen to square waves. You listen to music with all its content below 20KHz and anything above that doesn't matter.
The only non-ideal thing about music is that it's not band-limited because time matters - that's why compressing music with a lot of cymbals etc smears it across time.
This is why you low-pass audio before quantizing. If you put the square-wave through a low-pass filter before recording it digital and on high-quality tape, the digital recording will be more accurate than the tape.
This was a digital recording onto a professional digital system. It was put through a Nyquist filter.
Then again, is "accurate" in some mathematical sense the right term? Which one sounds more like music? Among audio professionals, there's a broad consensus that tape sounds better - enough so that there is a strong market for digital plug-ins that emulate the (mis)behavior of tape.
If a pure-digital master can emulate tape, then isn't it superior to tape?
(I know that many analog systems have soft-rolloff allowing you more leeway in setting your headroom, but modern ADCs have so much dynamic range that it's a poor operator who doesn't allocate themselves sufficient headroom).
I realize you weren't talking about 0-60 time, it was an analogy I introduced. Not meant to be technically similar, only a similar fraction of the whole. I think "driv[ing] in a straight line at 60mph" is a smaller fraction of car performance than frequency response of headphone performance, and 0-60 time more similar.
Actually, this is kind of a thing: speed measurement and reliability for cruise control. We can get close enough to be useful and quite comfortable. However, most cars don't directly measure their own speed anymore. Instead, it's inferred from other measurements.
"Sine wave measurements of audio gear ignore impulse response, intermodulation distortion, phase shift, and a host of other real-world physical device responses to real-world musical signals." Plus discussion of dynamic range; I don't think they were tested at different volumes.
Note that the article make no claim to be determining which headphones "sound better when stuck in human hears listening to music". And the article draws some specific conclusions and points out some of it's known limitations.
Humour me - what are some/any of the "bunch of things that should be measured" that were not, for a study titled "No correlation between headphone frequency response and retail price"?
I notice they also failed to test for "soundstage" "musicality" "warmth", or waterproofness, taste, and availability in a range of colours to suit your decor. Another "bunch of things" they didnt set out to test nor made any claims about.
It's obvious from the title that frequency response is being used as a proxy for quality, and it's outright stated multiple times such as "assuming that the perceived audio quality is largely determined by the spectral magnitude response of headphones".
But it's a bad way to test frequency response. And don't pull out those weird undefined words when this was a discussion of important objective audio-quality measures being ignored.
I don't think it is obvious. I have headphones I use for mixing, because of the frequency response. I don't use them to listen to music. Quality depends on what you are measuring, and no one claims a perfect frequency response necessarily means it sounds good. Often the opposite, since we use those types of headphones to find the flaws. So to ME, it's obvious that this is NOT about being a proxy for quality, and I'm inclined to believe that it is about what it claims to be about.
A sine sweep is still a very narrow and incomplete way to measure frequency response.
> it's obvious that this is NOT about being a proxy for quality
"Interestingly, sound quality does not seem to be a major attribute for purchase decisions."
"Root-mean square errors (RMSEs) were calculated across frequency for each headphone with respect to an assumed target curve to assess an objective quality metric."
"assuming that the perceived audio quality is largely determined by the spectral magnitude response of headphones"
I'm not going to disagree with you on the technical details, since I'm not knowledgeable enough in the area. But I'm going to take issue with one thing:
"Scientific, reductionist thinking is inadequate to get an accurate picture of the factors that matter to human listeners"
I think you're noting the difference between good and bad science -- science is certainly capable of putting together a picture of the factors that matter to human listeners. It may not have done so, but it's not a failure of the scientific process, it's a failure of the study(ies) involved.
I'm swinging to subjectivism not because I'm buying audio industry fluff (frankly, the self-styled objectivists are far more guilty of that), but rather because it at least covers the observed data. I'm listening with my ears.
Making up important-sounding quasi-science "objective" data (like the ever-popular THD) is an industry marketing ploy. People are insecure and want the "best", so pretend-science lets them think they're buying "best", rather than actually listening and judging subjectively, which is scary and hard and full of weird biases.
> Making up important-sounding quasi-science "objective" data (like the ever-popular THD) is an industry marketing ploy. People are insecure and want the "best", so pretend-science lets them think they're buying "best", rather than actually listening and judging subjectively, which is scary and hard and full of weird biases.
To be fair, the far more common trap is when people believe they're getting a better experience via the price-quality relationship. Or due to false authenticity. Or general brand marketing, etc.
Often these things fail the sniff test, and people who swear their $500 headphones are noticeably better will pick the cheaper headphone in a blind test.
The entire "audiophile" industry is more quasi-quality than quasi-science, though there's a healthy dose of whatever it takes to make the people with $ spend $.
That's true too, good points. And let's face it, this whole endeavor is about entertainment and artistry, so the subjectivist viewpoint is totally valid!
The factors that matter to human listeners is likely to be subjective. It seems likely that even such things as fidelity are likely to be subjective.
Which is to say, the scientific method isn't capable of answering all questions. It doesn't appear to be able to answer why I prefer this artist while disliking that artist.
I'm not even sure if things like preferences for art can be quantified. I'm all ears, if you've a way to study this.
I think a lot of art preferences can be quantified (kind of like how frequency response is actually an okay proxy for audio quality in headphones). My gripe is taking this common-sense case and trying to turn it into a universal theory of everything, dismissing actual human experience.
As an artist, I firmly believe we can at least quasi-objectively say "good art", or "bad art". However, this does not correlate well to personal preference. My love of White Castle sliders in no way suggests that they are actually good burgers. I can dislike a work of art and still know it's good.
Popularity can be a good proxy, too. Any hit song is a good song, even if you hate it.
Do you think the article is guilty of "trying to turn it into a universal theory of everything, dismissing actual human experience" anywhere? It's very specific in it's conclusions and discusses several of your objections in it's limitations section (at least from the Abstract - perhaps they're reaching much further in the full paper?)
"I think you're noting the difference between good and bad science ..."
Perhaps what he meant was "scientific thinking which is reductionist is inadequate ..."
His continued responses in this thread do not sound like the same old magical-audio-snake-oil ... but then again I have no expertise (or even experience) in this area.
I... don't know what you're talking about. Apply a log scale to linear data and it too represents a log scale.
> Recorded music has 8+ octaves of frequency range (the bottom octave plus a little extra is almost always rolled off in real-world recordings, to ease stress on downstream components that can't reproduce such low frequencies anyway), and 20-50db of useable dynamic range.
So? We all know that listening to an instrument live sounds "different" from a recording, and it's up to us to figure out how to improve both recording and listening fidelity.
Now an argument can be made that because audio formats today tend to have caps on fidelity that headphones only be measured against the maximum fidelity that the recording produces, but that is neither the argument that you are making nor is this a weakness in the data presented by the paper.
> Frequency response and total harmonic distortion aren't measured in these cases because they're useful or relevant. They're measured because they're easy to measure.
Yes and what should we measure? Phase graphs? Third order components? You're not making a meaningful argument here.
> The article's claim that one headphone can be easily morphed into another through mere equalization is, frankly, bullshit.
First of all this paper used novel methods to back the original assertion made in [1]. Also, let me quote the exact line from the conclusion:
"PCA can account for 90% of the variance across all measured headphones with six eigenvectors. The first eigenvector is similar to published target responses, while the second eigenvector represents a global spectral tilt."
> Audio reproduction is incredibly complex and difficult stuff. Trying to isolate one factor and saying "That explains everything!" is bad thinking.
And so is your appeal to non-authority. The gist of your entire argument is that linear sine sweeps are useless, and therefore the entire validity of the paper is moot and your non-scientific opinion is now superior.
The frequency response and the impulse response are correlated. You can just get one from the other through fourier transformation. Therefore it's perfectly fine to measure whatever is easier to measure, which is frequency response through sine waves for audio equipment. Phase shift can also be measured through this approach (and is required for calculating impulse reponse).
Sure, nonlinear distortions are not covered and also sure that the measurements won't cover everything which will influence the perceived sound. But nevertheless frequency response measurements are a very solid tool which allows objective comparions between various systems.
In theory, theory and practice are the same. In practice, however...
Yes, mathematically, you can get impulse response via Fourier transform. In practice, you're using a current to push a coil against a magnet into a spring. A real current, a real magnet, a real spring. You get all sorts of non-linear behaviors. On top of that, the spring (that is, the driver) retains impulse energy as mechanical energy, and then emits it back in a nonlinear way, by moving that coil in the magnetic field - turning a motor into a generator, producing seriously nonlinear signal back into the amplifier that produced the original signal, where it's picked up by the distortion-correcting negative feedback loop, and...
Yeah, that's not a simple Fourier transform anymore. If you had enough data points, it could be modeled mathematically. We don't have all the data, and don't yet have a good way to measure all that stuff.
That's really valid only in the case when the frequency response and the impulse response completely characterize a system--that is, only for linear and time-invariant systems. I know you covered this base by mentioning nonlinear distortions, but I think that's the big issue. Otherwise you could just correct the frequency response with equalization (assuming the frequency response isn't literally 0 ever).
For example, consider a system that just squares or cubes its input. The unit impulse response will look the same in both cases. The frequency response isn't even defined, since the steady state response to a sinusoid is not itself a sinusoid.
Came here to respond with something similar to the first paragraph. I was wondering if perhaps audio engineers mean something different when they're talking about impulse response, since my expertise is in signal processing for electromagnetic signals (vs. audio).
Yeah, I think we do. When audio engineers talk about impulse response, we're really talking about how accurately components can reproduce sharp-attack sounds. Snare drums, acoustic guitar strings, some synth sounds - these are examples of fast impulse response.
Microphones really struggle with this. So do mic preamps, recording media, amplifiers, and especially speakers. Recorded instruments sound wildly different from live-in-a-room instruments.
edit: I should add here that fast impulse response isn't necessarily what we want. Consider the common use of the relatively sluggish Shure SM57 on snare drums, rather than the much faster response of a small diaphragm condenser. The condenser usually just sounds harsh. The SM57 smooths out the sound. This is good, because nobody in their right mind actually sticks their ear one inch from a snare drum.
No, you're not talking about different things. You're talking about exactly the same thing in two domains: a signal which is transduced from air pressure to voltage in real time, and then later goes back in the other direction. Audio engineers should know that, and many of them do.
"Recorded instruments sound wildly different from live-in-a-room instruments." Yes, they do.
In the best case, you have two nearly perfectly accurate transducers located inside the not-quite-ear-canals of a dummy head doing the recording, and you play those two signals back to transducers located similarly in the head of your listener. This is the best case for accurate reproduction of instruments. You can get a recording that sounds nearly the same as being in the room where it happened.
In the most common "purist" method, recording engineers are using two or three mics arrayed in free space somewhere near where the audience would be, and hoping to capture the sensation of being in the room. This is problematic because the instruments are complex, moving three-dimensional shapes producing sound in a field which interacts with all sorts of things before being sampled by those 2-3 points. That throws away nearly all the available information. Then playback emanates from two not-really-point sources in a completely different room.
In the normal non-purist methods, engineers close-mike each of the instruments, take direct input from some electronic instruments, and may or may not try to capture some room ambience. Then they process everything to the point where they are as much a performer as the recorded players, and work on it over and over again until (hopefully) everyone is happy with it as a work of art. There should be no pretense that this is going to get you an accurate rendition of the feeling of being there, though.
Good points. Purist recording methods are almost unused in popular music; they're for classical or solo instruments only. There are few rational musical situations where an acoustic guitar is as loud as a drum kit, but we hear it all the time on records. Likewise, sounds aren't panned hard left or hard right, we don't record in giant churches, etc. Not to mention multiple overdubs by the same performer...
I don't really care if a recording is accurate. I care if it's enjoyable.
When I think of "the feeling of being there", I don't think of the feeling of being in the room with the musicians. I think of the feeling of what you're doing when you experience that music - where does it take you? I'll never be able to separate Pink Floyd's The Wall from making out with the girlfriend who introduced me to Pink Floyd, for example. Music has strong sense-memory, almost as strong as smell. That's what I want to engage - I want to make music you remember.
It's sometimes amazing to hear the albums which didn't have all that processing work done at the end. I love the sound of the DMB Lillywhite sessions which was leaked after recoding but not turned into an official album. I much prefer those versions to the later official ones.
> The frequency response and the impulse response are correlated. You can just get one from the other through fourier transformation.
It's true. It should be added, however, that when the frequency response is complex (and it always is with transducers) and the measurement conditions are not very repeatable, it's probably beneficial to also get an impulse response.
> Phase shift can also be measured through this approach
I thought headphones are zero-phase-shift devices? This is because of the very small volume of air between the membrane and the eardrum - they both move in lockstep, hence zero phase shift.
Human beings are remarkably good at compensating for deviations in frequency response. You can very quickly acclimatise to deviations much greater than +-10dB. The main shortcoming of mid-priced headphones is transient response, which we are acutely sensitive to. Heavy, flexible and poorly-damped drivers in highly resonant enclosures might look reasonably accurate in the frequency domain, but they have absolutely dismal real-world performance.
Audio engineers routinely choose transient response over frequency response, hence the popularity of Yamaha NS10 and Auratone 5C monitors in professional environments.
That is an awesome article! And this quote from it I find rings true relative to my criticism of the OP article...
"Why have I included a frequency-response curve here? I mentioned earlier that the frequency-response curves in a sales brochure are typically meaningless in terms of providing information that's useful to an end user. Actually, though, I'd go further than that, and suggest that in many respects making any judgment about the worth or likely value of a monitor by examining its frequency-response curve is not far short of pointless."
His science-based analysis of why the NS-10 is successful and popular is very enlightening, and a much better example of how to do audio science, imho. I learned a lot. I really wonder how it compares to the Tannoys that I use? My biggest gripe with the Tannoys is that they're too nice. In particular, they have a very sweet, non-fatiguing upper midrange. I worry that nasty things are sneaking by me.
I'm not an expert - I was into car audio back in the late 90's so some things got remembered.. But one thing about sine-wave frequency sweeps was that you'd have standing waves from interaction with the car environment (the famous Honda Civic "buzzing trunk" scenario). So they're really only good for first-pass level setting, and you should quickly go onto real music.
It's always helpful to listen to your mixes on a very familiar system, regardless of (or particularly due to) its inaccuracies. You have to remember that your music will be listened on almost everything - from alarm clock radios over default iPhone earplugs to the occasional proper system.
> It's like looking in the wrong place, because the light is better there.
Yes, and actually one of the challenges physicists face, is that looking by the lamp is not self evidently reasonable for most people. Of course you start with the easy case and then you build on that foundation, otherwise your analysis just hangs in thin air. (Plus you have to deal with the easy case anyhow if you want to do comprehensive work, you can as well start were you have a chance.)
I once heard a story in audiophile community, I don't recall the particularities, though. Guy went to audio exhibition and there was one booth with curtains covering the actual product. This is a new WIP product, we want to hear feedback without you actually seeing what it is to get a bit more objective response. Audience listened to some music playing from behind the curtains, discussed various strengths and weaknesses of the audio system. After a while curtains were raised to reveal live acoustic band playing.
Instruments themselves are imperfect, there is long chain of devices (recording, mixing, playback) until the signal reaches speakers (and then room until signal reaches ears), at higher quality levels it is quite often the case of how well all the inaccuracies of the system interact with each other, rather than how accurate certain components of the system are.
>> Frequency response and total harmonic distortion aren't measured in these cases because they're useful or relevant. They're measured because they're easy to measure.
We've evolved our senses around the signal components the pay off the best: the "easy" ones. Thus it isn't surprising that "subjective quality is mostly correlated with linear (spectral) attributes instead of non-linear (distortion) metrics," a claim supported by peer reviewed papers that have vastly greater credibility than you and your anecdotal vitriol.
They're doing a comparative study. To a large extent, it doesn't need to reflect reality. It just needs to be good enough to make a determination within the given sample.
> It's like measuring a car's performance by how well it can drive in a straight line at 60mph.
No, it's like comparing several cars and their MRSP based on their 0-60 performance, which is certainly a reasonable thing to do. Just because they didn't measure stopping distance, it doesn't invalidate the entire comparison.
Question: when checking the reference output, do you sit in the Subaru, or do you just have the speakers in a room with you? I would imagine the acoustics are fairly different.
Mixing music is very, very difficult. You have to deal with the inaccuracy of your equipment, the inaccuracy of your ears, and the ability to drift into boiled-frog mode, making something terrible that sounds good to you because you got here gradually.
Another kind of related tool I use extensively while mixing is reference recordings - listening to someone else's music on the same system, in alternation with what I'm mixing, to make sure I'm not boiling my frog.
But yes, listening to other reproduction systems - especially bad ones, like my Subaru speakers and older iPod earbuds - is critical. It helps me insure that what sounds good on my multi-thousand $$ studio monitors in a decent room sounds good everywhere else, too.
Another question: if it comes down to a choice of making something sound worse on consumer hardware, vs. sounding worse on studio hardware, you'll favor the consumer listening experience at the expense of the monitor output, right?
So, does this mean that it's a bad idea to use studio-grade equipment for personal listening of commercially-released (and so, presumably, mastered for consumer equipment) tracks?
And, if so, are alternate mixes of the same tracks that are sometimes released any better? A "club mix", for example (assuming you don't care about the changes to the track itself)?
And, one more question—given that mastering somewhat distorts the pre-master for a particular listening profile... is there any way for someone who wants to sample a track in their own production, to get access to the pre-master copy of the track they want to sample from, so they don't "accumulate mastering artifacts" the way that JPEG-recompression accumulates compression artifacts? I've seen, every once in a while, a band that has a "remix competition" release e.g. a Garage-band project with all the raw tracks embedded. Is it common to make a deal with a producer to get access to something like this on a one-off basis? Is it even technically feasible, or is pre-master track data lost/discarded after completion of mastering often enough that such requests don't make sense?
(Sorry, I never realized there were so many of these questions on my mind. Is there a sound-engineer Quora?)
Listen on whatever you like! I feel it's the mix/master engineers' jobs to create a final-product recording that sounds good in as many different environments as possible. Just my opinion, of course, and there are certainly recordings that will only sound good on fine hi-fi gear.
I like to think of a recording as a miniature of a real-world musical experience. I love cranking up my old Mesa Mark I guitar amp to the threshold of pain and wailing on it. It's an amazing sonic experience. But I cannot reproduce that sound on a pair of $20 Skullcandy earbuds. So instead of going for accuracy of reproduction, I'm trying to reproduce the feel, the vibe. Kind of like how a painting represents a landscape differently than a photograph, and neither really represents the landscape well. I'll take the painting, when it's truer to the feeling of the landscape.
This can be seen in dynamic range. Human hearing has tremendous dynamic range, and live acoustic instruments do as well. But your average modern record has no more than 20db, probably far less (often less than 6db). The "loudness war" of RMS vs peak volume is at play here, but more importantly, it plays to the dynamic weakness of consumer audio gear. Less dynamic range sounds "better", up to a point.
You don't want to listen to un-mastered mixes. You really don't.
It's common to use a boombox, Auratones (or similar), and Yamaha NS10Ms for this purpose, but cars do have their own audio quirks that make driving tests a good thing to do once you're close to finalizing the mix.
I'm trying to understand what "Log sine sweeps rather than linear sine sweeps were employed to allow verification that non-linear distortion components were virtually absent" actually means.
I understand that a log sine sweep means that the set of frequencies that are tested is itself logarithmic; e.g. you'd check 20Hz, 200Hz, 2KHz, 20KHz instead of 5Hz, 4005Hz, 8005Hz, etc. That makes sense to me. But I don't understand why log sine sweeps would allow you to check for non-linear distortion?
> The two headphones I rely on the most (Beyerdynamic DT880 and AKG K240) sound wildly different.
Former Psych major here whose studies specialized in perception (audio and visual)... Without validation using the scientific gold standard of a double-blind listening test, this is merely a belief that can easily be attributed to confirmation bias (see: popularity of Beats headphones)
Why and when would you pick the 880's and when the 240's? I have the 880's and they're pretty good so far, I compared them to a few AKG's in the store and I liked them a lot better for mixing.
Comfort and fatigue. I can listen to the 240s for longer, and they sound subjectively "flatter" to me. I actually like the sound of the 880s better. Plus the 880s are almost total isolation, which is really useful when tracking in the studio, or when working in a cubicle.
Some systems reproduce waveforms with a small delay that varies by frequency--i.e. high frequencies are slightly delayed relative to low ones, or vice versa, so the frequencies are out of phase with each other, relative to the original. Phase isn't intrinsically audible (though the brain does use it in stereo imaging) but it could be indirectly perceivable when it interacts with other aspects of a system, especially nonlinear aspects.
Why do you trust the OP's claims? How about all the people disagreeing with the OP?
Don't be so quick to think the top post is always right. The top comment of many HN threads are usually critical and sensationally written, and the child comments are usually in argument against the parent.
Good advice, even though in this case it's totally wrong, because I am absolutely right and not to be questioned. :)
To generalize more, those critical/sensational top comments are often written by experts with a very different perspective on the subject (as was mine), and most of the child comments are ignorant, angry howling about someone's sacred cow getting gored. On the other hand, a good child comment (and there are several here) can be really enlightening.
This attitude might have served you well this time, but who's to say every other time will be the same? I detect a far too eager appetite to embrace the brisk dismissals of others. It won't serve you well in the long term.
There are many articles posted here it will improve your very personality to read. Read them and don't make excuses not to. There will always be someone saying "XYZ article is bullshit". Many people here are actually pretty smart, and more than a hundred of them felt otherwise. Make your own decision.
Me neither of course. And an upvote is not an ironclad endorsement. But at least you look at what you upvote - or at least I hope you do.
You're not exactly disagreeing with me, are you?
The stupid thing is I actually happen to agree with "beat" that this study is basically bullshit. I just can't condone the attitude of the first replier. It worked this time but he or she has learned the wrong heuristic IMO. I guess I don't know how to express myself well but I think it's the wrong lesson to learn from criticism of the articles we see here - even if the criticism is (currently) right.
> You're not exactly disagreeing with me, are you?
I definitely disagree with "Many people here are actually pretty smart, and more than a hundred of them felt otherwise." An article upvote is not necessarily a statement of "this is not bullshit". Most of those people have likely not performed a critical analysis and decided on that issue. And there are no article downvotes, so there's always the possibility of an invisible majority against the article's conclusion.
And so exactly how is the top-most comment more reliable than the article then? Same people voting, same lack of critical analysis. At least the bloody author probably spent more than 5 seconds making up his mind.
> I don't even go to articles anymore
You're endorsing this as best practise. Right? Because that was my argument, that this is not best practise. Note that if everyone followed this practise, this site would fall apart.
edited - turns out i do want to continue arguing about it
It's good to read the articles, but that's not the part I objected to in your comment. I objected to the idea that there is always someone claiming bullshit and it's white noise, or that upvotes imply not-bullshit.
In other words, you can get a good idea of whether something is bullshit without reading, but you should still give it a look anyway. I agree with part of what you said, but not the other part.
Fair enough. On the face of it I don't disagree with your points. I should have expressed myself more clearly from the beginning.
Wish you'd look at it from the other direction though. Skipping reading articles because of assertions of the currently-top comment's dismissals is a recipe for groupthink and circle-jerks. Upvotes at least reflect a general consensus that an article has some merit. The top comment may indeed credibly refute an article - but the attitude of "I don't even go to articles anymore" is not that of someone seeking the truth. Instead, it's that of a lost soul looking for his "team" to join, and avoiding any contradictory information.
They do justify their methodology, I didn't think it sounded that unreasonable.
They flat out say that out of the two easily-measured factors, distortion and linear response, linear response correlates the most with subjective measures of audio reproduction according to prior research. I don't think they said that they had developed a foolproof methodology for absolutely determining the subjective reproductive quality of a headphone.
I think your description of the performance of headphones is based on soft, unscientific nonsense. Sure, there's more to headphones than single frequency response curves, but frequency response between the ear canal and the headphone is the only differing factor in audio reproduction quality.
If anything is flawed with the methodology, it would surely be with the lack of broad spectral testing or something equivalent. The fundamental characteristics of the driver are the diaphragm geometry, the mass of the driver, and the resistance of the suspension. The suspension changes, probably not linearly, with temperature. Frequency sweeps completely miss the point that the movement of a headphone driver is linear actuation, not some mystical frequency-domain process.
Oh, it's soft, unscientific nonsense. It is also direct experience as a semi-pro in the field.
Subjective experience is not science, but neither is it irrelevant. If someone's "science" does not explain observed subjective experience very well, then it shouldn't pretend that it does - and it really is not grounds to dismiss the subjective experience of experts by saying "BUT THOSE DUDES HAVE NUMBERS AND STUFF!!!!"
"Subjective experience is not science, but neither is it irrelevant."
Ah, anecdata.
Look, if you want to point to well controlled studies, etc, that say "these factors do not correlate well with subjective experience", that's awesome.
I'd then agree 100% "whatever we are measuring doesn't matter, we should measure something else".
But in your rage, you are conflating two issues here, and they shouldn't be conflated at all:
1. Was this study science, and properly performed science?
All available info seems to point to "yes, it was"
There is no reason for you to put air quotes around science, etc.
They set out what they are trying to measure and why: "This study quantifies variability of measured headphone response patterns and aims to uncover any correlations between headphone type, retail price, and frequency response."
They did not set out to say whether that has any bearing on subject experience.
In fact, they point out "The preferred response however seems to be listener, content, and headphone dependent <cite omitted>"
2. Does the thing they measured matter in any way to the subjective experience in the world?
You vehemently suggest "no".
I'm going to suggest if you want to convince people the answer is no, you should point them to data that says "the thing they measured, properly, doesn't matter", and not appeal to anecdote and authority.
They cite at least three studies thinking it matters:
"In particular, research suggests that the frequency (magnitude) response is a major factor in listener preference scores (Olive and Welti, 2012; Fleischmann et al., 2012; Olive et al., 2013)"
(and they are super careful not to suggest that listener preference scores completely correlate with subject experience)
but at the same time, admit
"Research suggests that factors influencing consumers' choice as to which model to purchase are mostly based on wireless functionality (Iyer and Jelisejeva, 2016) and attributes such as shape, design, and comfort (Jensen et al., 2016)."
They also admit the studies usually are small and that the body of work is not huge.
So, from my perspective, i feel like they are doing a fairly reasonable job of trying to present a relatively objective perspective on whether this matters or not.
I dunno. I think headphones track price-to-quality pretty well in the $20-250 range. Above that, it starts turning into luxury/status symbol stuff. This is my completely subjective opinion. What isn't opinion is that increases in quality are usually a matter of diminishing returns. It becomes increasingly expensive to get increasingly small incremental improvements.
Additionally, I think "sound quality" in headphones is very subjective. There are fine quality headphones that I really, really dislike (Grados, for example). I find a lot of expensive hi-fi headphones overly bright, too.
A good price/performance example is two headphones I keep around... my Beyerdynamic 880s, which I love, and the Sennheiser HD280. The 880s cost about twice as much, and fill the same role for me - closed-ear phones with very strong isolation, for performers tracking vocals or instruments. They can put a loud backing signal into the ear with only minimal leakage into the microphones, and easily block out other loud instruments in the same room. But the 880 sounds far better. It's flatter and more detailed. It's also much more comfortable for extended wear, better built, and more repairable. That's the $100 to $200 difference. But $200 to $400? Smaller change.
Fundamentally, there is nothing in the sound of a headphone other than specific spectral response (for a given head fit and ear). If you have an "ideal" headphone, you can emulate any headphone which is less than ideal.
If well-recorded music sounds bad on an ideal headphone, the recording is set up for a non-ideal headphone.
You'll notice in those response curves, there are common characteristics in the response of almost all the headphones they tested, within a fairly small margin (considering). Maybe for a mastering technician, the right thing to do is master for the average suboptimal headphone, and not for a linear response listening instrument; but that doesn't mean a headphone closer to the ideal is wrong, it just means that you need to filter audio that wasn't intended for its response profile.
There may not be a measurable criteria that drives price in all cases. Headphones are manufactured and priced for wildly different reasons. For instance, Beats are expensive because they're a status symbol. It has little to do with sound quality.
If you can't back your accusations here up with math, instead of handwavy "but it's so complicated and human ears are just, like, different, maaaaan", you're in no place to call the article bullshit.
This is signals and math, nothing more complicated. Sorry if that means that your prejudices are bullshit--but that's how most audiophile stuff goes.
Beat is criticising the specific (and sole) measure used in the article, and has provided a list of specific measurable properties that also have bearing on the issue of sound quality. I think your criticism is unfounded.
The complicated, handwavy stuff occurs when you try to map the measurements on to what people think is good. Maybe some day we'll have a good enough handle on that (psychoacoustic perceptual models, and so on) to call it just "signals and math", but right now it isn't.
We have excellent knowledge of psychoacoustics; it's why audio compression works so well. It's pretty straightforward actually. Psychovisual tuning, that's the one we're not good at.
Give me enough data points, and I could explain it all with the math.
We don't have enough data points. At a certain point, we perhaps can't have enough data points, just because the interactions are so complex.
Try reading Dekker's Drift into Failure. It's about failure analysis in complex systems (and why reductionist thinking is often a bad idea when trying to understand such failures), but it certainly applies to trying to "explain" the audible behavior of real-world sound reproduction with mere math.
edit: As a for-example... a speaker driver (like a headphone) is basically an electric motor attached to a spring (the diaphragm suspension). The suspension (spring) holds it at a zero point, and the motor moves it from the zero point, pushing air in the process. An electric motor consists of an AC-charged coil moving against a magnetic field. Now, if you look the other direction, a coil moving inside a magnetic field is an alternator, generating AC power.
So when the signal from the amplifier drives the motor that moves the speaker driver, energy gets stored in the spring - and then released back into the alternator, and pushed back into the terminals of the amplifier. That back signal is subject to serious nonlinearities from the suspension, including distortion, frequency response variations, and frequency-dependent group delay and phase shifts.
Most - but not all - of the back current from the speaker is absorbed by the output devices from the amp (which have high but not infinite impedance). What gets through is then picked up by the global negative feedback loop that is supposed to keep the amplifier linear, injecting it as phase-reversed signal into the input. Um.
This has a number of effects. First and foremost, it makes the amp/speaker interface much more sonically colored that it seems on the surface. Second, it blows up amplifiers when under enough strain - this is a real-world effect that any PA engineer has observed.
But go on, tell me again how my objections are just unscientific mystical hand-waving.
And with that, this study is bullshit.
Human beings don't listen to linear sine sweeps. We listen to music. Recorded music has 8+ octaves of frequency range (the bottom octave plus a little extra is almost always rolled off in real-world recordings, to ease stress on downstream components that can't reproduce such low frequencies anyway), and 20-50db of useable dynamic range.
Sine wave measurements of audio gear ignore impulse response, intermodulation distortion, phase shift, and a host of other real-world physical device responses to real-world musical signals. Scientific, reductionist thinking is inadequate to get an accurate picture of the factors that matter to human listeners.
Frequency response and total harmonic distortion aren't measured in these cases because they're useful or relevant. They're measured because they're easy to measure. It's like looking in the wrong place, because the light is better there. And the results? It's like measuring a car's performance by how well it can drive in a straight line at 60mph. Acceleration, braking, and turning are too hard to measure, so we ignore them...
I'm a musician and record producer. I've engineered and produced numerous albums, and rely on multiple different types of headphones for different purposes. The article's claim that one headphone can be easily morphed into another through mere equalization is, frankly, bullshit. The two headphones I rely on the most (Beyerdynamic DT880 and AKG K240) sound wildly different. Neither is "accurate". Neither are the Tannoy System 12 DMT midfield studio monitors I use for mixing, or the stock Subaru car speakers I use for reference to check the mixes from the Tannoys.
Audio reproduction is incredibly complex and difficult stuff. Trying to isolate one factor and saying "That explains everything!" is bad thinking.