ThomasXia
Member
- Joined
- Mar 6, 2019
- Messages
- 39
- Likes
- 100
Hey everyone, another forward of the designer's responses. This time a video was made in detail to critique the accuracy of NFS. I personally think it deserves more discussions.
【闲聊Klippel NFS与消声室测量】
Translation of the whole content with the help of DeepL:
Hello, everyone, this is GFSG119. I'm very sorry, because the content of this video is destined to be long and stinky. For those who don't know much about acoustic testing and are not interested, it may be difficult for them to listen, and for those who know a little bit, they may have some information and deep-rooted ideas that will make it difficult for them to accept what is said in this episode.
What is the topic of this video? It's about Klippel's near field scanner, whether it works as advertised or not, and whether it is better than the anechoic chamber test or not. To be noticed, this video has nothing to do with the X5 test ASR did at a few days ago. In fact, friends who have been following me for a long time will know that the content of my video has been mentioned one after another a year or two ago, and this video is just a more comprehensive and complete introduction. As I said in the first few states, ASR's test on the X5 itself has no obvious problem, but the X5 speakers did not leave the factory with such test results, and others did not test out such results. Only during the last six months some bizarre changes may have been produced. I don't want to say too much more about this matter. And to be honest, no matter what, just let others to evaluate the product. In fact, I have made more than one product that has been reviewed by ASR, including some other media, but these things are products that I have made in some previous companies, including car speakers that have been reviewed more. I do not want to say anything too much about these reviews themselves.
Back to the subject of this video, I'll take two points from Klippel's advertising or marketing story on the description of the anechoic chamber. From there, I'll start today's video. One of the points is that the anechoic chamber is not really an absolute ideal free field, it also has certain reflections. However, Klippel's test instrument can filter out this interference to get more accurate low frequency data.
Does such a test error exist in the actual anechoic chamber or not? From a certain point of view, yes. For example, if the anechoic chamber is small, or if there is a problem with the design or quality of the chamber itself, there will be some obvious error in the first frequency test. But if the anechoic chamber is a good quality, large anechoic chamber, it will not produce significant error in the low frequency test, or the frequency of the error can be called the cutoff frequency, it will become lower, there are more effective intervals, you can get a valid measurement value. Therefore, it is biased to use this reason to show that the anechoic chamber must have test errors in low frequencies.
It is obvious to generalize the shortcomings of some smaller or problematic anechoic chambers as the shortcomings of all anechoic chambers. What's more, this problem itself is relatively easy to judge. Because if there is really a certain residual reflection in the dealer's room, it will usually be shown in the test curve, that is, there will be some fluctuations in the test curve similar to the same speaker in the room test. Or such a defect, is easier to identify or easier to predict through theory.
Like Apple, Microsoft's anechoic chamber, or even the anechoic chamber of Guoguang in China, its test accuracy for a portion of the low frequency band must be higher than the accuracy calculated by an NFS in an ordinary room in the low frequency. I will explain this point later in the explanation of the NFS. In summary, the allowable operating range and cutoff frequency of an anechoic chamber can be pre-calculated and calibrated by measurement, rather than generalizing that all anechoic chambers have an order of magnitude problem, which is a rather uncritical statement. Another issue is that large temperature changes may have an effect on phase testing at high frequencies. First of all, there is something called air-conditioning, so I think this problem becomes quite naive. some people will say that in some test signal methods, the frequency response curve of high frequency will also be affected by the phase envelope
due to temperature changes, which in turn leads to errors. But in fact, this is only part of the test method. There are about half a dozen ways to measure the frequency response curve. Some tests like step sweep are not affected by complex envelopes. There are other test methods that pre-calibrate the test information. It can be understood that these test methods are inherently error-correcting.
Let's step back here and assume that the temperature difference is large and really causes errors in the test system. For a given test system, all the errors introduced by temperature are actually determined, and the differences caused by temperature can be compensated off with a pre-calibration. In fact, there are many acoustic tests with much higher temperatures than those just described. For example, the acoustic testing of aircraft rockets. For this kinds of civil product testing of speakers, when there is an error, the consequence is a slight deviation in tone at most. For the aircraft rocket, if the test has such a large error and there is no solution, it is really going to blow up. So this idea is again a very obvious generalization of some special cases.
And in fact, the NFS itself is usually tested over a long period of time, such as seven or eight hours or more. If the room where the device is located is not temperature-controlled, it is likely that the temperature in the room will change by more than the 2 degrees that he tells in his marketing story. That is, NFS's own tests also have to correct and compensate for errors. If the air-conditioning is not on and if there is no temperature-controlled test, then the difference in temperature over time must be even greater.
Above is a short description of the problem of the anechoic chamber. The next step is to talk about the NFS. Of course I have talked about some of them one after another before. Some people will think I don't know much about NFS, but I've actually been using NFS much earlier than some of the test media, and have done a lot of testing and verification. And having worked for professional testing companies like audio precision and Gras, I have reason to believe that I know more about testing instruments like NFS than some of the testing media.
Here is a short introduction to the Klippel NFS test principle. Simply put, a mechanical arm is used to scan the speaker at a relatively close distance, through the algorithm to calculate the acoustic performance of the speaker at any point in space. This test principle in a simplified theoretical model can achieve a more desirable test results. But the problem is that the traditional anechoic chamber just described, even if there are errors are caused by practical factors. In fact, any real-world test instrument is subject to error, which I believe is well understood.
What does a theoretically feasible NFS look like in reality if it has more error than the anechoic chamber? There are many problems with NFS itself, two main points here, one is the robotic arm and the other is the realistic algorithm problem or software problem. In the description before, you should understand that the NFS test requires the microphone to move to hundreds or thousands of points in the coordinate system. The microphone itself is not moved by the force from vaccum, but by a robotic arm. The robotic arm used by NFS does not actually meet the requirements of the algorithm itself. Because the arm of NFS is not an ideal two-dimensional straight line, it is a solid with a volume. This causes the microphone angle to deviate from the center of the sound at some positions during the movement of the arm in the r and φ axes. Because the robot arm used in NFS is not a multi-axis industrial robot like this one, there is no way to adjust the angle arbitrarily with such a simple robot arm.
On the other hand, the microphone itself is also directional. When testing data at a certain point, the microphone should maintain an angle that points to the center of the sound. So if the angle does not point to the center of the sound or if it changes, then the test value of the microphone is not the true frequency curve of the point. The bigger problem is the z-axis, which is the vertical direction. Because the NFS arm can only rotate in the horizontal plane, there is no pitch angle. Therefore, the vertical direction of motion will vary in all heights except for the plane where the sound center is located, and the angle between the microphone and the sound center will make the original error at each point of the test.
I draw such a diagram here, I believe many people should be able to see. So actually, the problem is over here. Because raw data have errors, in the subsequent calculation the credibility of the results is necessarily not high. The error is correlated with the specific size of the box, the ratio of the length, width and height, as well as the specific point of the test settings. I personally speculate that this is also one of the reasons why the final output of the NFS sometimes has errors at high frequencies, and why some times there is a large error, and some times there is a small error.
As a comparison, here is a look at one of Gras's array microphones. All microphones are distributed or their trajectories are on a sphere and the angle of each position is pointing in the direction of the sound center. This is such an obvious problem with the professional base. Although it seems that when I asked one of the review media who uses NFS, he didn't even notice, Klippel himself must have known. So he wrote this in the corner of one of the NFS data sheets in small print. This is the equivalent of him admitting that test errors do exist. And Klippel said here that even if a 1/4-inch microphone is used, in practice it can only mitigate this error but not eliminate it.
And the microphone mentioned here is a point where Klippel blew himself up. Because the test microphone used by NFS is actually a product of my previous company, and the microphone that comes with NFS by default, its accuracy is only plus or minus 2 DB. Taking into account the off-axis pointing error mentioned earlier, the error is probably plus or minus 3 DB.
Klippel says in the data sheet that their test instrument can achieve plus or minus 0.1 DB accuracy. But how does it do this with plus or minus 0.1 DB accuracy? Combined with some of the data I will list later, I can assert here that Klippel's accuracy for the test instrument must be inflated. I take responsibility for this conclusion.
In short, the NFS arm is an oversimplification and a new variable in itself. For example, the arm or the mechanical structure itself may resonate, or in some cases the arm may simply jam or not move smoothly. Because the one-dimensional mechanical structure is not an air-floating guide and does not have any lubrication.
The second major or bigger problem is the software algorithm of NFS. Theoretically this algorithm works, but the algorithm needs to be programmed and validated before actual uses. I believe that many people who are not in the acoustics business but in the software business should be aware of this. We cannot equate a theoretical algorithm with the final performance of the software.
First of all, the software is inherently buggy, such as some times it will crash and flash or suddenly stop.
Forget about these problems for the moment. The first major problem of the software itself is the fitting error of low and high frequencies. The frequency response curve of the NFS output is calculated, it will have a judgment itself. Putting aside the validity of the NFS fitting algorithm, and assuming its own judgment is correct, many times the error is beyond its own allowable range in both high and low frequency fitting. In other words, the NFS itself admits that the values output from these bands are invalid. Here's where things go wrong.
At the beginning of the video, Klippel's marketing will say that the anechoic chamber is not valid. Of course, as we have just explained, this is an act of generalization. But what kind of behavior is it when NFS itself has errors and doesn't mention them, and I use my ideal values to compare with your bad actual values? And more critically, even if the test data meets Klippel's own judgment of the fit criteria, the test results may still be in error. Because these people's reasoning is that for anechoic chambers the low frequencies are not accurate.
Of course, as at the very beginning of the video I also introduced, despite the low frequency and even high frequency they said the phase is not accurate at middle frequencies. And why is there deviation in the middle frequencies as well? For example, I gave an example here, this speaker in the NFS test is this result.
But in the anechoic chamber with Soundcheck measurement is another result. Sorry about not importing different test instrument results to the same comparison box, it is rather troublesome. I'm lazy, so you can pause to see which frequency bands deviate from each other.
Further, I'm using another model of Klippel's audio analyzer here. For this speaker, in Klippel's another test instrument and in the NFS test, the results are also different. These two examples are not exceptional, even a random example just show this deviation. Ignoring low-frequency and high-frequency deviations, there still exist deviations positioned in the middle frequency part. And if you look carefully, you should be able to find that the frequency band where they deviate from is also different.
So here I have reason to suspect that the deviation is an error in the NFS's own test results, rather than an error in the anechoic chamber measurements. Because it is neither low frequency nor high frequency. And I can get similar results in different anechoic chambers with different test instruments. The deeper reason is that the test in the dealer's room is the actual measured value, while the frequency response curve derived from the NFS is a calculated value, a predicted value, not a measured value.
For your understanding, I will show a video clip here.
(Video clip from the face-to-face program on the CCTV news channel, the topic is "Ding Zhongli: What is a fair abatement program")
Reporter Chai Jing:If it is simulated and calculated, all this is credible, is not also a basis?
Academician Ding Zhongli: How do you know it is credible?
Reporter Chai Jing:We almost believe in all the data in the lab.
Academician Ding Zhongli: It is not a laboratory, it is a computer, how do you know it is credible or not credible?
Reporter Chai Jing: Ding academician, of course we know that there are opposing and skeptical voices in the scientific community, but we have the impression that because the IPCC is such a research organization, it is also scientists from various countries together to come up with a report. And also because there is a report as a basis, countries all over the world will go to a climate conference, so we have the impression that he has the approval of the mainstream scientific community.
Academician Ding Zhongli: Is there a mainstream of scientists?
Reporter Chai Jing:We understand the mainstream is ......
Academician Ding Zhongli: Should the mainstream be based on the number of people on a side? Science is the judgment of truth.
Some measurement media will say that there exist well-matched results between NFS and the anechoic chamber test, or some NFS tests are very flat results. First of all, this is a paradox in itself, because according to these media, since the anechoic chamber is not accurate, if the NFS and the anechoic chamber test results are consistent, doesn't it mean that the NFS is also not accurate?
And if the NFS test results are straight, does that necessarily mean that the test results are accurate? As you can see from my example, if there is a discrepancy between the anechoic chamber and the NFS calculation, I think it is entirely possible that the NFS calculation is flat. But the anechoic chamber or the real example of not flat, for example, I just put Newman's KH120 speaker with an EQ, so that it looks flat in the NFS results. Even so, there may be a greater bias in the anechoic chamber.
I believe Klippel himself should not say such an amateur argument, because whether it is to judge the measured value or calculated value, can not simply use such an example as the basis, especially there are biases and conflicting some results.
Some of my own tests of professional amplified sound boxes also have examples of NFS calculations and anechoic chamber tests that match almost perfectly. But I can not ignore the existence of deviations because of these ideal results, and I can not use some ideal results to conclude that the NFS calculation results must be accurate, especially in the absence of any comparison. If you don't understand, I'll give you an inappropriate example.
The value calculated by NFS in this way is like some almanac that some NPC characters in the game in the science fiction "The Three Body Problem I, Remembrance of Earth's Past" think they have, and the actual measurement comes out with some deviation. It is the equivalent of the flying star phenomenon in the "Three body" game, and actually has a much higher probability than the flying star in the "Three body" world. I cannot be selective and ignore these flying star phenomena just because I want to show that the almanac is free of deviations. Nor can I go so far as to assume that because this almanac did predict the pattern of the three-body motion at a certain time, the subsequent three-body motion must also follow this almanac.
I cannot pretend to be a great prophet just because I predicted a partially correct result. Besides, in the previous talkings about the robot arm, I have already said that there is error in the raw data input, plus the algorithm itself may have some problems mentioned earlier, and the negative gets positive, and getting a relatively correct conclusion may only be a chance event, while the error is inevitable.
Back to the accuracy of the test instrumentation and calibration of the judgment. Real-world test instruments are not Christmas toys, you say it is accurate or not, is to follow the basic law. Usually a test instrument, its precision or accuracy is to be calibrated or verified according to such a system.
Products like Gras and AP have gone through a rigorous system of stating their accuracy, rather than me saying I'm accurate and I don't have any comparisons or any higher level of standards to calibrate against, not like NFS.
In my own software, I show that my fit is valid, I output the valid value, so my output at this time is calculated out of the frequency curve is no error. It's more accurate than the anechoic chamber. How can it look like this! A test instrument, its accuracy needs to be calibrated and verified by a calibration unit that is one level higher than its accuracy. And if you think you're more accurate than everyone else, you need to be certified by the International Conference on Weights and Measures, not by yourself saying that you're accurate.
And here I've seen some people who will sophomorically argue that the Gras microphone used on the NFS is calibrated, or that it is a sensitivity calibration. The calibration result cannot be equated to the deviation between the final calculated frequency response curve of the NFS output and the standard value. These are two completely different concepts. What we have just described is the calibration of the sensor, while here we are going to use the output of the final calculated frequency response curve of the NFS, which is calibrated and verified between this result and some other more standard metric facilities. There are also international standards for the reliability and consistency of test instruments. As I said earlier, the AP and Gras test facilities have these well established, I can show that I do meet the standards of the accuracy that I claim, or the process. If NFS also has a more rigorous calibration process, I think that process or procedure should be made public. And for a system that outputs calculated values, even like Sean Olive's speaker prediction models, there are confidence intervals, there are confidence levels, and there are fitting errors.
And for example, in this graph, you cannot use the fact that there are predicted and measured values that fall perfectly on the curve to prove that this prediction method itself is free of any error, especially if there are other points that deviate from the prediction method. If you use only the points on the fitted curve and ignore the other points with deviations, you are actually trying to prove that this conclusion is accurate. And proving that this conclusion is accurate is just wrong in the underlying logic.
Note that the fit error I am referring to here is not the fit error shown in the NFS, but rather the comparison between the calculated values output by the NFS test instrument and the measured values relative to a standard anechoic chamber, or a system that can be used for calibration. Are there any deviations between the calculated values of the NFS and the actual measured values? If so, what is the approximate deviation? What is the tolerance range?
Instead of test results like this showing resonances that do not exist and are not measured in the anechoic chamber. Clearly, the test results here alone show that the tolerance range claimed by Klippel is exceeded.
The NFS sometimes tests for resonances that do not exist. This video is long enough, and there are many more questions, if I get a chance to talk about them later.
For example, there are times when even in the pre-test calibration of the device is the signal-to-noise ratio are no problem, but after the final test is completed, it shows itself that the fitting error is not acceptable at all, even if this test is performed in a normal room. This explains the claim by Klippel that they can algorithmically filter out interference in the room. Such a claim is not valid, at least not every time. Either that or it is clear that such calculations are done, but there are problems shown in the fitting error. In fact, if you just look at the curve alone, especially the low and midrange part, you can see that it is actually no different from the normal speaker. So if the fitting error itself is problematic, I actually do not know if it is right or wrong to calculate the value at this time if there is no comparison.
Perhaps some people will also say, you produced so smelly and long a video, why not directly speaking with Klippel? You're talking about this in public here, are you targeting someone or this company now? In fact, all the problems I mentioned here and even more problems not mentioned, I have reflected with Klippel more than once before. For example, two different NFSs, they measured the same sample, and there was a 2 DB deviation in low frequency. But when this sample was tested in three different anechoic chambers, the results basically all overlapped. I am not alone in my experience with the product itself, but because it was a project in my company's R&D, I am not comfortable speaking directly and publicly on this occasion. But I've talked to Klippel about these issues, and even approached their after-sales service in 2019, but many of the issues were never resolved. And for example, many of the problems I talk about in this video are not just found by me, and I didn't make the video to raise the bar.
So at the end of the video, I'll give a few comments that I think are constructive. First of all, NFS needs to be compared and verified with a better anechoic chamber when testing, and it is better if the results are similar. If the results are different, we need to find out what is causing the difference and who is accurate and who is not. The second point, if their results are similar, NFS can roughly achieve the effect of the anechoic chamber. Simply speaking is no problem. But if you must be serious and say that the NFS must be more accurate than the anechoic chamber, this conclusion is not valid. The third suggestion is that I think Klippel should improve its own robot arm and improve its own algorithm to solve these problems and make these kinds of bugs less, instead of avoiding or claiming that I have no problem with this stuff. In fact, such a video is not only stink and long, but also offending. Why didn't the people who found these problems before talk about them publicly like I did? After all, no one wants to offend such a highly respected old professor. And to discuss this issue, I suggest that it is best to discuss the principles of acoustics and the principles of measurement, rather than which person has said what. Because if you simply use a certain authority said no problem, directly because Klippel invented, so there must be no problem, but also to discuss these technologies why?
The original content in Chinese:
哈喽,大家好,我是鬼斧神工119。非常抱歉,因为这期视频所讲的内容注定是又臭又长的。因为对声学测试不了解,不感兴趣的人可能很难听得进去,而对于一些了解那么一点点的人,可能他们之前接受到的一些信息,有一些根深蒂固的观念,会导致他们很难再去接受这期内容所讲的东西。
这一期视频我要讲的主题是什么?主要就是关于 Klippel 的 near field scanner 测试仪器本身,它到底有没有宣传上的这样的效果,以及和消声室测试对比到底孰优孰劣。不,不过这里事先要声明,这期视频所讲的内容和我前几天 X5 在 ASR 这里测试是没有什么关系的。其实关注我比较久的朋友会知道,我这期视频所讲的内容在一两年前都有陆续的提到过,这期视频只是做一个比较综合完整的介绍。正如我前几条状态所说的, ASR 关于 X5 的测试本身是没有什么明显问题的,但是我出厂的时候不是这样的测试结果,别人测出来也不是这样的结果,只能是中间的这半年可能产生了一些诡异的变化。这个事情我不想再说太多。而且说实话,不论如何,做产品就是要让别人去评价的。其实我所做的产品也不止一款之前被 ASR 测评过,包括其他的一些媒体,只不过这些东西是我在之前的一些公司所做的产品,包括车载音箱被测评的就更多了。测评这件事情本身我是不想发表什么太多意见的。
回这期视频的主题,我这里就从 Klippel 的广告或者营销故事中两点对消声室的描述。从这里为切入点来开启今天的视频。其中一点是,消声室其实并不是绝对的理想自由场,它也存在一定的反射。 Klippel 的测试仪器可以滤除掉这样的干扰,得到更精确的低频数据。
实际的消声室中到底存不存在这样的测试误差?从某种角度上是存在的。比如如果消声室很小,又或者消声室本身的设计或者质量存在一定问题,在第一频测试中会存在一定比较明显的误差。但是如果当消声室是一个质量比较好,体积比较大这样的消声室,它在低频的测试中就不会产生比较明显的误差,或者产生误差的频率也可以被称作截止频率,就会变得更低,就有更多有效的区间,可以获得有效的测量值。所以,如果以此理由来说明消声室在低频中一定存在测试误差,这样是有失偏颇的。
这种把一部分较小或者存在一些问题的消声室的缺点说成是所有消声室的缺点,明显是以偏概全的。更何况这个问题本身是比较容易判断的。因为如果销商室内真的存在一定的残余反射,通常会体现在测试曲线中,也就是测试曲线中会存在一些类似于音箱在房间中测试的波动。或者这样的缺陷,是比较容易识别或者比较容易通过理论去预测的。
像苹果、微软的消声室,哪怕是国光的消声室,它对于低频一部分频段的测试精度一定是高于一台 NFS 在一个普通房间中低频所计算出来的精度。稍后我会在对 NFS 的解释中说明这一点。总之,消声室允许工作的范围和截止频率是可以预先计算,并且可以通过测量进行标定的,而不是泛泛的说所有的消声室都存在一个量级的问题,这种说法是相当不严谨的。另一个问题是,温度变化较大可能会对高频的相位测试产生影响。首先,有一种东西叫做空调,所以我认为这个问题就显得相当的naive。有些人会说,在有些测试信号方法中,高频的频响曲线也会受到温度变化而导致的相位包络的影响,进而导致误差。但实际上这只是一部分测试方法。频响曲线的测量大概有六七种方法。有些像 step sweep 这样的测试就不会受到复杂的包络的干扰。还有其他一些预先标定测试信息的测试方法。可以理解为这些测试方法本身就是具有纠错能力的。
我们这里退一步讲,假设温度差异较大,而且真的引起了测试系统的误差。对于给定的测试系统,所有温度引入的误差其实都是确定的,只要进行预先的标定,就可以补偿掉温度带来的差异。其实声学测试中有很多温度远远高于刚才所说的这些测试。比如飞机火箭的声学测试。音箱这种民用产品测试,如果有误差,顶多也就是音色稍微有点偏差,但是飞机火箭如果测试有这么大的误差,还没有解决方法,是真的要炸的。所以这又是一个很明显的以偏概全的说法。
而且事实上, NFS 自身的测试通常要经历很长的时间,比如七八个小时甚至更久。如如果设备所在的房间没有温控设施,房间内变化的温度很有可能高于他在营销故事中所讲的 2 度。也就是 NFS 自己的测试也要纠错和补偿。如果不开空调,如果没有温控测试,那么长时间所带来的温度差异一定会更大。
简短的说明了消声室的问题。接下来就说一下 NFS 的问题。当然我之前就陆续讲过一些。有些人会认为我对 NFS 不了解,但实际上我用 NFS 比一些测试媒体都要早很多,做过非常多的测试和验证。而且我在像 audio precision 和 Gras 这样的专业测试公司工作过,所以我有理由相信我比某些测评媒体更了解像 NFS 这样的测试仪器。
这里首先简短介绍一下 Klippel NFS 的测试原理。简单来说就是采用一个机械臂对音箱在较近的距离进行扫描,通过算法算出这个音箱在空间中任意一点的声学表现。这个测试原理在简化的理论模型中可以实现比较理想的测试效果。但是问题在于,刚才所讲的传统的消声室即便存在误差也是实际因素导致的。事实上,任何现实世界中的测试仪器都是存在误差的,这一点我相信大家很好理解。
理论上可行的 NFS 在现实中是什么样?如果它的误差比消声室还要大。 NFS 自身的问题有很多,这里主要两点,一个是机械臂,一个是现实中的算法问题或者软件问题。在刚才的描述中,大家应该明白, NFS 的测试需要麦克风在坐标系中运动几百或者几千个点。麦克风本身不是靠原力凭空移动的,而是要靠机械臂。这里就引入了一系列新的变量,这些变量都有可能导致更多的误差。NFS 所采用的机械臂其实并不符合它算法本身的要求。因为NFS 的机械臂并不是理想的二维直线,它是一个实体,是存在体积的。这一方面会导致机械臂在 r 轴和 φ 轴的运动过程中,有些位置麦克风的角度会偏离与声中心的连线。因为 NFS 所采用的机械臂不是像这样的多轴的工业机器人,这种比较简单的机械臂没有办法做到任意角度的调节。
而另一方面,麦克风本身也是存在指向性的。麦克风在测试某个点的数据时,应该保持指向声中心的角度。所以一旦这个角度没有指向声中心或者产生了变化,那么这个时候麦克风的测试值就不是这个点真实的频向曲线。而问题更大的是 z 轴,也就是竖直方向。因为 NFS 的机械臂只能在水平面内转动,是没有俯仰角的。所以竖直方向的运动除了声中心所在的平面以外,其他所有的高度都会因为麦克风与声中心的夹角变化,而而使得每个点的测试存在原始误差。
我这里画这样一个示意图,相信很多人应该能看明白。所以其实到这里这个问题已经结束了。 raw data 都是存在误差的,存在错误的后续计算的可信度必然是不高的。而且误差和箱体的具体尺寸、长宽高比例以及测试设置的具体的点都有关。我个人推测这也是 NFS 最终输出结果在高频有时存在误差的原因之一,以及为什么有些时候存在较大的误差,有些时候存在较小的误差的原因之一。
作为对比,这里可以给大家看一下 Gras 的一个阵列麦克风。所有麦克风都是分布或者它们的轨迹是在一个球面上的,并且每一个位置的角度都指向声中心的方向。这个就是专业基础这么明显的问题。虽然我在问某个采用 NFS 的测评媒体的时候,他似乎根本就没有注意过,但是 Klippel 自己肯定是知道的。所以他在 NFS 的一个 data sheet 的角落里用一个小字写下了这样一段话。这也相当于他承认测试误差确实是存在的。而且克里普这里所说的即便采用了 1/ 4 英寸的麦克风,实际中也只能减轻这个误差而不能消除。
而这里提到的麦克风,算是 Klippel 自爆或者说露出黑胶的一个点。因为 NFS 所采用的测试麦克风实际上是我之前一家公司的产品,而 NFS 默认自带的麦克风,它的精度只有正负 2 DB。在考虑到刚才所说的指向性离轴这样的误差,可能这个误差在正负 3 DB。
而 Klippel 在 data sheet 中说,他们的测试仪器能够达到正负 0. 1 DB 的测试精度。它所采用的我们的传感器只有正负 2 DB 的精度,甚至可能是正负 3 DB。但是它又是怎样做到的正负 0. 1 DB 的精度呢?再结合我后面会列出的一些数据,我可以在这里断言 Klippel 对测试仪器的精度一定是虚标。我为这个结论负责。
总之, NFS 的机械臂过于简化,而且本身就是一个新的变量,比如机械臂或者机械结构本身就有可能产生共振,又或者有些时候机械臂会直接卡住或者移动不顺滑。因为一维的机械结构既不是气浮导轨,也没有任何润滑。第二个比较大或者更大的问题是 NFS 的软件算法。理论上一个算法可行,但算法落地是需要编程并且经过验证的。我相信很多即便不是做声学而是做软件的人应该都清楚。我们不能把一个理论上的算法等同于软件最终的实际表现。首先,这个软件本来就是存在 BUG 的,比如有些时候会死机闪退或者突然停住。暂且去掉这些问题。软件本身的第一个大问题是低频和高频的拟合误差。 NFS 的输出的频响曲线是计算出来的,它自身会有一个判断。先抛开 NFS 拟合判断的算法是否有效,假设它自身的判断是对的,很多时候,在高频和低频拟合,误差也会超出它自身允许的范围。换句话说, NFS 自己就相当于承认了这些频段所输出的值是无效的。这里就出了问题了。
在视频的开头, Klippel 的 marketing 会说消声室不准。当然我们刚才已经解释过了,是一种以偏概全的行为。但是 NFS 自己有误差却不提,我用我的理想值和你的不好的实际值进行对比,这是什么行为?而更关键的是,即便测试数据符合 Klippel 自身判断的拟合标准,测试结果也有可能存在误差。因为这些人的理由是消声室低频是不准的。当然,最开始的视频中我也介绍了,把低频甚至高频他们说的相位不准的频段也去掉。而为什么中频也会存在偏差?比如我这里举了一个例子,这个音箱在NFS 中测试是这样的结果,但是在消声室中用 Soundcheck 测是另外一个结果。这里比较抱歉。不同的测试仪器导入到相同的一个对比框中,确实比较麻烦。我比较偷懒,大家可以暂停去看它们之间哪些频段有偏差。
更进一步,我这里用的是 Klippel 的另一种型号的音频分析仪。这一支音箱在 Klippel 的另一种测试仪器与在 NFS 测试中,结果也是存在不同的。这两个例子并不是特例,这是我随便找出来的就发现存在偏差的。而且这里既不是低频偏差,我们也不管他们高频的偏差,在中频部分,他们也是存在偏差的。而且如果仔细看,应该可以发现这两支音箱它们偏差的频段也是不一样的。
所以这里我有理由去怀疑这个偏差是NFS 自身测试结果的误差,而不是消声室测量存在误差。因为这既不是低频,也不是高频。而且我可以在不同的消声室用不同的测试仪器测试到相似的结果。而更深层次的原因是因为在销商室中测试是实际的测量值,而 NFS 所导出的频响曲线是计算值,是预测值,不是测量值。方便大家理解,我这里播一段视频片段。
(视频片段来自CCTV新闻频道的面对面节目,丁仲礼:什么是公平的减排方案)
柴静记者:如果它模拟计算出来,这一切是可信,不也是一个依据吗?
丁仲礼院士:你怎么知道它可信?
柴静记者:我们几乎是信仰实验室里所有数据。
丁仲礼院士:它不是实验室,它是计算机,你怎么知道它可信还是不可信?
柴静记者:丁院士,我们当然知道科学界有反对和怀疑的声音,但是给我们的印象是,因为 IPCC 的这样的一个研究的组织,它也是各国的科学家在一起拿出一份报告。而且也是因为有报告做基础,全世界的国家会到去开一个气候的大会,所以给我们的印象他是得到了主流科学界的认同的。
丁仲礼院士:科学家有主流吗?
柴静记者:我们理解的主流是……
丁仲礼院士:科学家是根据人多人少来定的吗?科学是真理的判断。
有些测评媒体会说存在和消声室匹配很好的结果,或者一些 NFS 测试很平直的结果。首先这本身就是一个悖论,因为按照这些媒体的说法,既然消声室是不准的, NFS 如果和消声室测试结果一致,不就恰恰说明 NFS 也不准了吗?而 NFS 的测试结果平直,就一定能说明测试结果是准确的吗?从我刚才的例子中就能看到,如果消声室和 NFS 的计算结果存在偏差,我认为完全有可能存在 NFS 计算出来的结果是平直的。但是消声室中或者现实中不平直的例子,比如我刚才把纽曼的 KH120 的音箱加一个 EQ,让它在 NFS 的结果中看起来平直。即便这样,可能在消声室中存在更大的偏差。
我相信Klippel自己应该是不会说出这么业余的论据的,因为不论是判断测量值还是计算值,都不能简单的用这样的个例作为依据,尤其是存在有偏差有冲突的一些结果。我自己测试的一些专业扩声音箱也有 NFS 计算结果和消声室测试几乎完美匹配的例子。但我不能因为存在这些理想的结果,就忽视掉存在偏差的结果,更不能用一些理想的结果去得出 NFS 的计算结果一定准确的结论,尤其是在没有任何对比的情况下。如果大家还不明白,我这里举一个不恰当的例子。
NFS 这样计算出来的值就像是《三体》游戏中一些 NPC 角色所自认为的一些万年历,而实际的测量出来的一些有偏差。它的情况相当于《三体》游戏中的飞星现象,而且实际上概率远比三体世界中的飞星的概率要高得多。我不能因为我要说明万年历是没有偏差的,就有选择的,忽视掉这些飞星现象。也不能因为这本万年历在一定时间内确实预测到了三体运动的规律,就去认为之后的三体运动也一定会遵循这本万年历。
我不能因为预测了一部分正确的结果,就冒充伟大的先知。况且在机械臂的内容中,我就已经说了,输入的原始数据就存在误差,再加上刚才提到的算法本身可能存在一些问题,负得正,得到相对比较正确的结论也许才是偶然的事件,而错误是必然。回测试仪器准确性的判断和标定。现实世界中的测试仪器并不是圣诞节玩具,你说它准不准,是要遵循基本法的。通常一个测试仪器,它的精度或者准确性是要按照这样的一个体系去进行标定或者是验证的。像 Gras 和 AP 的产品都是经过了严格的这套体系去说明它的精度,而不是我自己说我精度很高,我没有任何的对比或者任何更高 level 的这种标准去进行标定,而不是像 NFS 这样的。在我自己的软件中,我显示我的拟合是有效的,我输出的有效值,所以我这个时候输出的计算出来频向曲线就是没有误差的。比消声室更准的。怎么能这个样子。一个测试仪器,它的精度需要比它精度更高一级这样的标定单位进行标定和验证。而如果你认为你比所有人都要精度更高,你要获得国际度量衡大会的认证,而不是你自己说自己很精确。
而且这里我还看到有一些人会诡辩, NFS 上面用的 Gras 麦克风是校准过的,或者是灵敏度的校准。校准结果不能等同于 NFS 最终计算输出的频响曲线与标准值之间的偏差。这完全是两个概念。刚才所说的是对传感器校准,而这里我们要用到的是 NFS 最终计算出来的频赏曲线的输出结果,是这个结果和其他一些更标准的度量设施之间进行标定和验证。而对于测试仪器的可靠性和一致性也有相关的国际标准。就像我刚才所说, AP 和 Gras 的测试设备有这些完善的,我可以拿得出来证明我确实是符合我所宣称的精度的这样的标准,或者说这样的过程。如果 NFS 也有这样的比较严谨的一个标定过程,我认为这个过程或者流程应该公开出来。而对于一个输出计算值的系统,哪怕像Sean Olive的音箱预测模型,也有置信区间,也有置信度,也有拟合误差。
而且比如这张图,你不能用有的预测值和实测值完美的落在曲线上,就说明这种预测方法本身是没有任何误差的,尤其是存在其他偏离预测方法的点。如果只用拟合曲线上的点,而忽视掉其他有偏差的点,实际上是为了证明这个结论准确。而证明这个结论准确,在底层逻辑上就是错误的。
注意,我这里说的拟合误差并不是 NFS 里面显示的拟合误差,而是 NFS 测试仪器输出的计算值和相对来比较标准的消声室,或者可以用于标定的系统的测量值之之间的对比。 NFS 的计算值与实际的测量值到底有没有偏差?如果有这个偏差大概是多少?这个公差范围是多少?而不是像这样的测试结果中显示出一些不存在的,在消声室中也测量不到的谐振。显然,仅从这里的测试结果也能够看出,已经超出了 Klippel 所宣称的公差范围。
NFS 有些时候就是会测试出一些本身不存在的谐振。这个视频说到这里已经很长了,还有非常非常多的问题,如果以后我有机会再讲。比如有些时候即便在设备前期测试校准的时候是信噪比都没有问题的,但是最后测试完成之后,它自身显示的拟合误差就是不可用的。而这就是在一个正常的房间中去测试的。这也就说明要 Klippel 所宣称的,他们可以通过算法滤除房间的干扰。这样的宣传是不成立的,至少不是每一次都行的。要么就说明明完成了这样的计算,但是拟合误差里显示却有问题。实际上如果只是单独看曲线,尤其是中低频的部分,大家可以看出其实它跟正常的音箱没有什么区别。所以如果拟合误差本身显示都是有问题的,我实际上如果在没有对比的情况下,根本就不知道这个时候它计算出来值到底是对的还是错的。
也许有些人也会说,你讲这么又臭又长的视频,为什么不直接跟 Klippel 去讲?你在这里公开讲这个的事情,是不是要针对他怎么样的?其实我所说的所有的问题,包括更多的问题,我之前都跟 Klippel 反映过不止一次。比如两台不同的NFS,他们测同一个样品,低频就有 2 DB 的偏差。但是这个样品在三个不同的消声室中测试,基本上都是重合的。产品本身也不是我一个人经历,只不过因为是之前公司研发时的项目,不方便直接公开在这个场合讲。但是我这些问题都跟 Klippel 反映过,甚至在 2019 年的时候,我就找过他们的售后,但是很多问题都是不了了之。再比如我这一期视频中所讲的很多问题,也不只是我一个人发现,我做视频也不是为了抬杠。
所以在视频的最后,我给出几个我认为有建设性的意见。首先第一点, NFS 在测试的时候需要和比较好的消声室进行对比和验证,如果结果比较相似比较好。如果结果是不同的,需要找出到底是哪里导致的不同,到底谁准谁不准。第二个点,如果他们的结果是相似的,可以粗略或者大体上的 NFS 能够达到消声室的效果,这样简单的来说是没问题的。但是如果你一定要较真说 NFS 一定比消声室更准,这个结论是不成立的。第三个建议是,我认为Klippel要么改进自己的机械臂,要么改善完善自己的算法,把这些问题解决,让这些 bug 变得更少,而不是回避或者一味的宣称我这个东西都是没问题的。其实讲这样一期视频,除了又臭又长以外,也得罪人。为什么之前发现这些问题的人都没有像我这样公开去讲呢?毕竟谁也不想得罪这样一位德高望重的老教授。而且要讨论这个问题,我建议最好还是从声学原理和测量的原理去讨论,而不是哪个人说了什么。因为如果单纯用某个权威的人说了没问题,直接说是因为 Klippel 发明了,所以肯定没问题,还讨论这些技术干嘛?
【闲聊Klippel NFS与消声室测量】
Translation of the whole content with the help of DeepL:
Hello, everyone, this is GFSG119. I'm very sorry, because the content of this video is destined to be long and stinky. For those who don't know much about acoustic testing and are not interested, it may be difficult for them to listen, and for those who know a little bit, they may have some information and deep-rooted ideas that will make it difficult for them to accept what is said in this episode.
What is the topic of this video? It's about Klippel's near field scanner, whether it works as advertised or not, and whether it is better than the anechoic chamber test or not. To be noticed, this video has nothing to do with the X5 test ASR did at a few days ago. In fact, friends who have been following me for a long time will know that the content of my video has been mentioned one after another a year or two ago, and this video is just a more comprehensive and complete introduction. As I said in the first few states, ASR's test on the X5 itself has no obvious problem, but the X5 speakers did not leave the factory with such test results, and others did not test out such results. Only during the last six months some bizarre changes may have been produced. I don't want to say too much more about this matter. And to be honest, no matter what, just let others to evaluate the product. In fact, I have made more than one product that has been reviewed by ASR, including some other media, but these things are products that I have made in some previous companies, including car speakers that have been reviewed more. I do not want to say anything too much about these reviews themselves.
Back to the subject of this video, I'll take two points from Klippel's advertising or marketing story on the description of the anechoic chamber. From there, I'll start today's video. One of the points is that the anechoic chamber is not really an absolute ideal free field, it also has certain reflections. However, Klippel's test instrument can filter out this interference to get more accurate low frequency data.
Does such a test error exist in the actual anechoic chamber or not? From a certain point of view, yes. For example, if the anechoic chamber is small, or if there is a problem with the design or quality of the chamber itself, there will be some obvious error in the first frequency test. But if the anechoic chamber is a good quality, large anechoic chamber, it will not produce significant error in the low frequency test, or the frequency of the error can be called the cutoff frequency, it will become lower, there are more effective intervals, you can get a valid measurement value. Therefore, it is biased to use this reason to show that the anechoic chamber must have test errors in low frequencies.
It is obvious to generalize the shortcomings of some smaller or problematic anechoic chambers as the shortcomings of all anechoic chambers. What's more, this problem itself is relatively easy to judge. Because if there is really a certain residual reflection in the dealer's room, it will usually be shown in the test curve, that is, there will be some fluctuations in the test curve similar to the same speaker in the room test. Or such a defect, is easier to identify or easier to predict through theory.
Like Apple, Microsoft's anechoic chamber, or even the anechoic chamber of Guoguang in China, its test accuracy for a portion of the low frequency band must be higher than the accuracy calculated by an NFS in an ordinary room in the low frequency. I will explain this point later in the explanation of the NFS. In summary, the allowable operating range and cutoff frequency of an anechoic chamber can be pre-calculated and calibrated by measurement, rather than generalizing that all anechoic chambers have an order of magnitude problem, which is a rather uncritical statement. Another issue is that large temperature changes may have an effect on phase testing at high frequencies. First of all, there is something called air-conditioning, so I think this problem becomes quite naive. some people will say that in some test signal methods, the frequency response curve of high frequency will also be affected by the phase envelope
due to temperature changes, which in turn leads to errors. But in fact, this is only part of the test method. There are about half a dozen ways to measure the frequency response curve. Some tests like step sweep are not affected by complex envelopes. There are other test methods that pre-calibrate the test information. It can be understood that these test methods are inherently error-correcting.
Let's step back here and assume that the temperature difference is large and really causes errors in the test system. For a given test system, all the errors introduced by temperature are actually determined, and the differences caused by temperature can be compensated off with a pre-calibration. In fact, there are many acoustic tests with much higher temperatures than those just described. For example, the acoustic testing of aircraft rockets. For this kinds of civil product testing of speakers, when there is an error, the consequence is a slight deviation in tone at most. For the aircraft rocket, if the test has such a large error and there is no solution, it is really going to blow up. So this idea is again a very obvious generalization of some special cases.
And in fact, the NFS itself is usually tested over a long period of time, such as seven or eight hours or more. If the room where the device is located is not temperature-controlled, it is likely that the temperature in the room will change by more than the 2 degrees that he tells in his marketing story. That is, NFS's own tests also have to correct and compensate for errors. If the air-conditioning is not on and if there is no temperature-controlled test, then the difference in temperature over time must be even greater.
Above is a short description of the problem of the anechoic chamber. The next step is to talk about the NFS. Of course I have talked about some of them one after another before. Some people will think I don't know much about NFS, but I've actually been using NFS much earlier than some of the test media, and have done a lot of testing and verification. And having worked for professional testing companies like audio precision and Gras, I have reason to believe that I know more about testing instruments like NFS than some of the testing media.
Here is a short introduction to the Klippel NFS test principle. Simply put, a mechanical arm is used to scan the speaker at a relatively close distance, through the algorithm to calculate the acoustic performance of the speaker at any point in space. This test principle in a simplified theoretical model can achieve a more desirable test results. But the problem is that the traditional anechoic chamber just described, even if there are errors are caused by practical factors. In fact, any real-world test instrument is subject to error, which I believe is well understood.
What does a theoretically feasible NFS look like in reality if it has more error than the anechoic chamber? There are many problems with NFS itself, two main points here, one is the robotic arm and the other is the realistic algorithm problem or software problem. In the description before, you should understand that the NFS test requires the microphone to move to hundreds or thousands of points in the coordinate system. The microphone itself is not moved by the force from vaccum, but by a robotic arm. The robotic arm used by NFS does not actually meet the requirements of the algorithm itself. Because the arm of NFS is not an ideal two-dimensional straight line, it is a solid with a volume. This causes the microphone angle to deviate from the center of the sound at some positions during the movement of the arm in the r and φ axes. Because the robot arm used in NFS is not a multi-axis industrial robot like this one, there is no way to adjust the angle arbitrarily with such a simple robot arm.
On the other hand, the microphone itself is also directional. When testing data at a certain point, the microphone should maintain an angle that points to the center of the sound. So if the angle does not point to the center of the sound or if it changes, then the test value of the microphone is not the true frequency curve of the point. The bigger problem is the z-axis, which is the vertical direction. Because the NFS arm can only rotate in the horizontal plane, there is no pitch angle. Therefore, the vertical direction of motion will vary in all heights except for the plane where the sound center is located, and the angle between the microphone and the sound center will make the original error at each point of the test.
I draw such a diagram here, I believe many people should be able to see. So actually, the problem is over here. Because raw data have errors, in the subsequent calculation the credibility of the results is necessarily not high. The error is correlated with the specific size of the box, the ratio of the length, width and height, as well as the specific point of the test settings. I personally speculate that this is also one of the reasons why the final output of the NFS sometimes has errors at high frequencies, and why some times there is a large error, and some times there is a small error.
As a comparison, here is a look at one of Gras's array microphones. All microphones are distributed or their trajectories are on a sphere and the angle of each position is pointing in the direction of the sound center. This is such an obvious problem with the professional base. Although it seems that when I asked one of the review media who uses NFS, he didn't even notice, Klippel himself must have known. So he wrote this in the corner of one of the NFS data sheets in small print. This is the equivalent of him admitting that test errors do exist. And Klippel said here that even if a 1/4-inch microphone is used, in practice it can only mitigate this error but not eliminate it.
And the microphone mentioned here is a point where Klippel blew himself up. Because the test microphone used by NFS is actually a product of my previous company, and the microphone that comes with NFS by default, its accuracy is only plus or minus 2 DB. Taking into account the off-axis pointing error mentioned earlier, the error is probably plus or minus 3 DB.
Klippel says in the data sheet that their test instrument can achieve plus or minus 0.1 DB accuracy. But how does it do this with plus or minus 0.1 DB accuracy? Combined with some of the data I will list later, I can assert here that Klippel's accuracy for the test instrument must be inflated. I take responsibility for this conclusion.
In short, the NFS arm is an oversimplification and a new variable in itself. For example, the arm or the mechanical structure itself may resonate, or in some cases the arm may simply jam or not move smoothly. Because the one-dimensional mechanical structure is not an air-floating guide and does not have any lubrication.
The second major or bigger problem is the software algorithm of NFS. Theoretically this algorithm works, but the algorithm needs to be programmed and validated before actual uses. I believe that many people who are not in the acoustics business but in the software business should be aware of this. We cannot equate a theoretical algorithm with the final performance of the software.
First of all, the software is inherently buggy, such as some times it will crash and flash or suddenly stop.
Forget about these problems for the moment. The first major problem of the software itself is the fitting error of low and high frequencies. The frequency response curve of the NFS output is calculated, it will have a judgment itself. Putting aside the validity of the NFS fitting algorithm, and assuming its own judgment is correct, many times the error is beyond its own allowable range in both high and low frequency fitting. In other words, the NFS itself admits that the values output from these bands are invalid. Here's where things go wrong.
At the beginning of the video, Klippel's marketing will say that the anechoic chamber is not valid. Of course, as we have just explained, this is an act of generalization. But what kind of behavior is it when NFS itself has errors and doesn't mention them, and I use my ideal values to compare with your bad actual values? And more critically, even if the test data meets Klippel's own judgment of the fit criteria, the test results may still be in error. Because these people's reasoning is that for anechoic chambers the low frequencies are not accurate.
Of course, as at the very beginning of the video I also introduced, despite the low frequency and even high frequency they said the phase is not accurate at middle frequencies. And why is there deviation in the middle frequencies as well? For example, I gave an example here, this speaker in the NFS test is this result.
But in the anechoic chamber with Soundcheck measurement is another result. Sorry about not importing different test instrument results to the same comparison box, it is rather troublesome. I'm lazy, so you can pause to see which frequency bands deviate from each other.
Further, I'm using another model of Klippel's audio analyzer here. For this speaker, in Klippel's another test instrument and in the NFS test, the results are also different. These two examples are not exceptional, even a random example just show this deviation. Ignoring low-frequency and high-frequency deviations, there still exist deviations positioned in the middle frequency part. And if you look carefully, you should be able to find that the frequency band where they deviate from is also different.
So here I have reason to suspect that the deviation is an error in the NFS's own test results, rather than an error in the anechoic chamber measurements. Because it is neither low frequency nor high frequency. And I can get similar results in different anechoic chambers with different test instruments. The deeper reason is that the test in the dealer's room is the actual measured value, while the frequency response curve derived from the NFS is a calculated value, a predicted value, not a measured value.
For your understanding, I will show a video clip here.
(Video clip from the face-to-face program on the CCTV news channel, the topic is "Ding Zhongli: What is a fair abatement program")
Reporter Chai Jing:If it is simulated and calculated, all this is credible, is not also a basis?
Academician Ding Zhongli: How do you know it is credible?
Reporter Chai Jing:We almost believe in all the data in the lab.
Academician Ding Zhongli: It is not a laboratory, it is a computer, how do you know it is credible or not credible?
Reporter Chai Jing: Ding academician, of course we know that there are opposing and skeptical voices in the scientific community, but we have the impression that because the IPCC is such a research organization, it is also scientists from various countries together to come up with a report. And also because there is a report as a basis, countries all over the world will go to a climate conference, so we have the impression that he has the approval of the mainstream scientific community.
Academician Ding Zhongli: Is there a mainstream of scientists?
Reporter Chai Jing:We understand the mainstream is ......
Academician Ding Zhongli: Should the mainstream be based on the number of people on a side? Science is the judgment of truth.
Some measurement media will say that there exist well-matched results between NFS and the anechoic chamber test, or some NFS tests are very flat results. First of all, this is a paradox in itself, because according to these media, since the anechoic chamber is not accurate, if the NFS and the anechoic chamber test results are consistent, doesn't it mean that the NFS is also not accurate?
And if the NFS test results are straight, does that necessarily mean that the test results are accurate? As you can see from my example, if there is a discrepancy between the anechoic chamber and the NFS calculation, I think it is entirely possible that the NFS calculation is flat. But the anechoic chamber or the real example of not flat, for example, I just put Newman's KH120 speaker with an EQ, so that it looks flat in the NFS results. Even so, there may be a greater bias in the anechoic chamber.
I believe Klippel himself should not say such an amateur argument, because whether it is to judge the measured value or calculated value, can not simply use such an example as the basis, especially there are biases and conflicting some results.
Some of my own tests of professional amplified sound boxes also have examples of NFS calculations and anechoic chamber tests that match almost perfectly. But I can not ignore the existence of deviations because of these ideal results, and I can not use some ideal results to conclude that the NFS calculation results must be accurate, especially in the absence of any comparison. If you don't understand, I'll give you an inappropriate example.
The value calculated by NFS in this way is like some almanac that some NPC characters in the game in the science fiction "The Three Body Problem I, Remembrance of Earth's Past" think they have, and the actual measurement comes out with some deviation. It is the equivalent of the flying star phenomenon in the "Three body" game, and actually has a much higher probability than the flying star in the "Three body" world. I cannot be selective and ignore these flying star phenomena just because I want to show that the almanac is free of deviations. Nor can I go so far as to assume that because this almanac did predict the pattern of the three-body motion at a certain time, the subsequent three-body motion must also follow this almanac.
I cannot pretend to be a great prophet just because I predicted a partially correct result. Besides, in the previous talkings about the robot arm, I have already said that there is error in the raw data input, plus the algorithm itself may have some problems mentioned earlier, and the negative gets positive, and getting a relatively correct conclusion may only be a chance event, while the error is inevitable.
Back to the accuracy of the test instrumentation and calibration of the judgment. Real-world test instruments are not Christmas toys, you say it is accurate or not, is to follow the basic law. Usually a test instrument, its precision or accuracy is to be calibrated or verified according to such a system.
Products like Gras and AP have gone through a rigorous system of stating their accuracy, rather than me saying I'm accurate and I don't have any comparisons or any higher level of standards to calibrate against, not like NFS.
In my own software, I show that my fit is valid, I output the valid value, so my output at this time is calculated out of the frequency curve is no error. It's more accurate than the anechoic chamber. How can it look like this! A test instrument, its accuracy needs to be calibrated and verified by a calibration unit that is one level higher than its accuracy. And if you think you're more accurate than everyone else, you need to be certified by the International Conference on Weights and Measures, not by yourself saying that you're accurate.
And here I've seen some people who will sophomorically argue that the Gras microphone used on the NFS is calibrated, or that it is a sensitivity calibration. The calibration result cannot be equated to the deviation between the final calculated frequency response curve of the NFS output and the standard value. These are two completely different concepts. What we have just described is the calibration of the sensor, while here we are going to use the output of the final calculated frequency response curve of the NFS, which is calibrated and verified between this result and some other more standard metric facilities. There are also international standards for the reliability and consistency of test instruments. As I said earlier, the AP and Gras test facilities have these well established, I can show that I do meet the standards of the accuracy that I claim, or the process. If NFS also has a more rigorous calibration process, I think that process or procedure should be made public. And for a system that outputs calculated values, even like Sean Olive's speaker prediction models, there are confidence intervals, there are confidence levels, and there are fitting errors.
And for example, in this graph, you cannot use the fact that there are predicted and measured values that fall perfectly on the curve to prove that this prediction method itself is free of any error, especially if there are other points that deviate from the prediction method. If you use only the points on the fitted curve and ignore the other points with deviations, you are actually trying to prove that this conclusion is accurate. And proving that this conclusion is accurate is just wrong in the underlying logic.
Note that the fit error I am referring to here is not the fit error shown in the NFS, but rather the comparison between the calculated values output by the NFS test instrument and the measured values relative to a standard anechoic chamber, or a system that can be used for calibration. Are there any deviations between the calculated values of the NFS and the actual measured values? If so, what is the approximate deviation? What is the tolerance range?
Instead of test results like this showing resonances that do not exist and are not measured in the anechoic chamber. Clearly, the test results here alone show that the tolerance range claimed by Klippel is exceeded.
The NFS sometimes tests for resonances that do not exist. This video is long enough, and there are many more questions, if I get a chance to talk about them later.
For example, there are times when even in the pre-test calibration of the device is the signal-to-noise ratio are no problem, but after the final test is completed, it shows itself that the fitting error is not acceptable at all, even if this test is performed in a normal room. This explains the claim by Klippel that they can algorithmically filter out interference in the room. Such a claim is not valid, at least not every time. Either that or it is clear that such calculations are done, but there are problems shown in the fitting error. In fact, if you just look at the curve alone, especially the low and midrange part, you can see that it is actually no different from the normal speaker. So if the fitting error itself is problematic, I actually do not know if it is right or wrong to calculate the value at this time if there is no comparison.
Perhaps some people will also say, you produced so smelly and long a video, why not directly speaking with Klippel? You're talking about this in public here, are you targeting someone or this company now? In fact, all the problems I mentioned here and even more problems not mentioned, I have reflected with Klippel more than once before. For example, two different NFSs, they measured the same sample, and there was a 2 DB deviation in low frequency. But when this sample was tested in three different anechoic chambers, the results basically all overlapped. I am not alone in my experience with the product itself, but because it was a project in my company's R&D, I am not comfortable speaking directly and publicly on this occasion. But I've talked to Klippel about these issues, and even approached their after-sales service in 2019, but many of the issues were never resolved. And for example, many of the problems I talk about in this video are not just found by me, and I didn't make the video to raise the bar.
So at the end of the video, I'll give a few comments that I think are constructive. First of all, NFS needs to be compared and verified with a better anechoic chamber when testing, and it is better if the results are similar. If the results are different, we need to find out what is causing the difference and who is accurate and who is not. The second point, if their results are similar, NFS can roughly achieve the effect of the anechoic chamber. Simply speaking is no problem. But if you must be serious and say that the NFS must be more accurate than the anechoic chamber, this conclusion is not valid. The third suggestion is that I think Klippel should improve its own robot arm and improve its own algorithm to solve these problems and make these kinds of bugs less, instead of avoiding or claiming that I have no problem with this stuff. In fact, such a video is not only stink and long, but also offending. Why didn't the people who found these problems before talk about them publicly like I did? After all, no one wants to offend such a highly respected old professor. And to discuss this issue, I suggest that it is best to discuss the principles of acoustics and the principles of measurement, rather than which person has said what. Because if you simply use a certain authority said no problem, directly because Klippel invented, so there must be no problem, but also to discuss these technologies why?
The original content in Chinese:
哈喽,大家好,我是鬼斧神工119。非常抱歉,因为这期视频所讲的内容注定是又臭又长的。因为对声学测试不了解,不感兴趣的人可能很难听得进去,而对于一些了解那么一点点的人,可能他们之前接受到的一些信息,有一些根深蒂固的观念,会导致他们很难再去接受这期内容所讲的东西。
这一期视频我要讲的主题是什么?主要就是关于 Klippel 的 near field scanner 测试仪器本身,它到底有没有宣传上的这样的效果,以及和消声室测试对比到底孰优孰劣。不,不过这里事先要声明,这期视频所讲的内容和我前几天 X5 在 ASR 这里测试是没有什么关系的。其实关注我比较久的朋友会知道,我这期视频所讲的内容在一两年前都有陆续的提到过,这期视频只是做一个比较综合完整的介绍。正如我前几条状态所说的, ASR 关于 X5 的测试本身是没有什么明显问题的,但是我出厂的时候不是这样的测试结果,别人测出来也不是这样的结果,只能是中间的这半年可能产生了一些诡异的变化。这个事情我不想再说太多。而且说实话,不论如何,做产品就是要让别人去评价的。其实我所做的产品也不止一款之前被 ASR 测评过,包括其他的一些媒体,只不过这些东西是我在之前的一些公司所做的产品,包括车载音箱被测评的就更多了。测评这件事情本身我是不想发表什么太多意见的。
回这期视频的主题,我这里就从 Klippel 的广告或者营销故事中两点对消声室的描述。从这里为切入点来开启今天的视频。其中一点是,消声室其实并不是绝对的理想自由场,它也存在一定的反射。 Klippel 的测试仪器可以滤除掉这样的干扰,得到更精确的低频数据。
实际的消声室中到底存不存在这样的测试误差?从某种角度上是存在的。比如如果消声室很小,又或者消声室本身的设计或者质量存在一定问题,在第一频测试中会存在一定比较明显的误差。但是如果当消声室是一个质量比较好,体积比较大这样的消声室,它在低频的测试中就不会产生比较明显的误差,或者产生误差的频率也可以被称作截止频率,就会变得更低,就有更多有效的区间,可以获得有效的测量值。所以,如果以此理由来说明消声室在低频中一定存在测试误差,这样是有失偏颇的。
这种把一部分较小或者存在一些问题的消声室的缺点说成是所有消声室的缺点,明显是以偏概全的。更何况这个问题本身是比较容易判断的。因为如果销商室内真的存在一定的残余反射,通常会体现在测试曲线中,也就是测试曲线中会存在一些类似于音箱在房间中测试的波动。或者这样的缺陷,是比较容易识别或者比较容易通过理论去预测的。
像苹果、微软的消声室,哪怕是国光的消声室,它对于低频一部分频段的测试精度一定是高于一台 NFS 在一个普通房间中低频所计算出来的精度。稍后我会在对 NFS 的解释中说明这一点。总之,消声室允许工作的范围和截止频率是可以预先计算,并且可以通过测量进行标定的,而不是泛泛的说所有的消声室都存在一个量级的问题,这种说法是相当不严谨的。另一个问题是,温度变化较大可能会对高频的相位测试产生影响。首先,有一种东西叫做空调,所以我认为这个问题就显得相当的naive。有些人会说,在有些测试信号方法中,高频的频响曲线也会受到温度变化而导致的相位包络的影响,进而导致误差。但实际上这只是一部分测试方法。频响曲线的测量大概有六七种方法。有些像 step sweep 这样的测试就不会受到复杂的包络的干扰。还有其他一些预先标定测试信息的测试方法。可以理解为这些测试方法本身就是具有纠错能力的。
我们这里退一步讲,假设温度差异较大,而且真的引起了测试系统的误差。对于给定的测试系统,所有温度引入的误差其实都是确定的,只要进行预先的标定,就可以补偿掉温度带来的差异。其实声学测试中有很多温度远远高于刚才所说的这些测试。比如飞机火箭的声学测试。音箱这种民用产品测试,如果有误差,顶多也就是音色稍微有点偏差,但是飞机火箭如果测试有这么大的误差,还没有解决方法,是真的要炸的。所以这又是一个很明显的以偏概全的说法。
而且事实上, NFS 自身的测试通常要经历很长的时间,比如七八个小时甚至更久。如如果设备所在的房间没有温控设施,房间内变化的温度很有可能高于他在营销故事中所讲的 2 度。也就是 NFS 自己的测试也要纠错和补偿。如果不开空调,如果没有温控测试,那么长时间所带来的温度差异一定会更大。
简短的说明了消声室的问题。接下来就说一下 NFS 的问题。当然我之前就陆续讲过一些。有些人会认为我对 NFS 不了解,但实际上我用 NFS 比一些测试媒体都要早很多,做过非常多的测试和验证。而且我在像 audio precision 和 Gras 这样的专业测试公司工作过,所以我有理由相信我比某些测评媒体更了解像 NFS 这样的测试仪器。
这里首先简短介绍一下 Klippel NFS 的测试原理。简单来说就是采用一个机械臂对音箱在较近的距离进行扫描,通过算法算出这个音箱在空间中任意一点的声学表现。这个测试原理在简化的理论模型中可以实现比较理想的测试效果。但是问题在于,刚才所讲的传统的消声室即便存在误差也是实际因素导致的。事实上,任何现实世界中的测试仪器都是存在误差的,这一点我相信大家很好理解。
理论上可行的 NFS 在现实中是什么样?如果它的误差比消声室还要大。 NFS 自身的问题有很多,这里主要两点,一个是机械臂,一个是现实中的算法问题或者软件问题。在刚才的描述中,大家应该明白, NFS 的测试需要麦克风在坐标系中运动几百或者几千个点。麦克风本身不是靠原力凭空移动的,而是要靠机械臂。这里就引入了一系列新的变量,这些变量都有可能导致更多的误差。NFS 所采用的机械臂其实并不符合它算法本身的要求。因为NFS 的机械臂并不是理想的二维直线,它是一个实体,是存在体积的。这一方面会导致机械臂在 r 轴和 φ 轴的运动过程中,有些位置麦克风的角度会偏离与声中心的连线。因为 NFS 所采用的机械臂不是像这样的多轴的工业机器人,这种比较简单的机械臂没有办法做到任意角度的调节。
而另一方面,麦克风本身也是存在指向性的。麦克风在测试某个点的数据时,应该保持指向声中心的角度。所以一旦这个角度没有指向声中心或者产生了变化,那么这个时候麦克风的测试值就不是这个点真实的频向曲线。而问题更大的是 z 轴,也就是竖直方向。因为 NFS 的机械臂只能在水平面内转动,是没有俯仰角的。所以竖直方向的运动除了声中心所在的平面以外,其他所有的高度都会因为麦克风与声中心的夹角变化,而而使得每个点的测试存在原始误差。
我这里画这样一个示意图,相信很多人应该能看明白。所以其实到这里这个问题已经结束了。 raw data 都是存在误差的,存在错误的后续计算的可信度必然是不高的。而且误差和箱体的具体尺寸、长宽高比例以及测试设置的具体的点都有关。我个人推测这也是 NFS 最终输出结果在高频有时存在误差的原因之一,以及为什么有些时候存在较大的误差,有些时候存在较小的误差的原因之一。
作为对比,这里可以给大家看一下 Gras 的一个阵列麦克风。所有麦克风都是分布或者它们的轨迹是在一个球面上的,并且每一个位置的角度都指向声中心的方向。这个就是专业基础这么明显的问题。虽然我在问某个采用 NFS 的测评媒体的时候,他似乎根本就没有注意过,但是 Klippel 自己肯定是知道的。所以他在 NFS 的一个 data sheet 的角落里用一个小字写下了这样一段话。这也相当于他承认测试误差确实是存在的。而且克里普这里所说的即便采用了 1/ 4 英寸的麦克风,实际中也只能减轻这个误差而不能消除。
而这里提到的麦克风,算是 Klippel 自爆或者说露出黑胶的一个点。因为 NFS 所采用的测试麦克风实际上是我之前一家公司的产品,而 NFS 默认自带的麦克风,它的精度只有正负 2 DB。在考虑到刚才所说的指向性离轴这样的误差,可能这个误差在正负 3 DB。
而 Klippel 在 data sheet 中说,他们的测试仪器能够达到正负 0. 1 DB 的测试精度。它所采用的我们的传感器只有正负 2 DB 的精度,甚至可能是正负 3 DB。但是它又是怎样做到的正负 0. 1 DB 的精度呢?再结合我后面会列出的一些数据,我可以在这里断言 Klippel 对测试仪器的精度一定是虚标。我为这个结论负责。
总之, NFS 的机械臂过于简化,而且本身就是一个新的变量,比如机械臂或者机械结构本身就有可能产生共振,又或者有些时候机械臂会直接卡住或者移动不顺滑。因为一维的机械结构既不是气浮导轨,也没有任何润滑。第二个比较大或者更大的问题是 NFS 的软件算法。理论上一个算法可行,但算法落地是需要编程并且经过验证的。我相信很多即便不是做声学而是做软件的人应该都清楚。我们不能把一个理论上的算法等同于软件最终的实际表现。首先,这个软件本来就是存在 BUG 的,比如有些时候会死机闪退或者突然停住。暂且去掉这些问题。软件本身的第一个大问题是低频和高频的拟合误差。 NFS 的输出的频响曲线是计算出来的,它自身会有一个判断。先抛开 NFS 拟合判断的算法是否有效,假设它自身的判断是对的,很多时候,在高频和低频拟合,误差也会超出它自身允许的范围。换句话说, NFS 自己就相当于承认了这些频段所输出的值是无效的。这里就出了问题了。
在视频的开头, Klippel 的 marketing 会说消声室不准。当然我们刚才已经解释过了,是一种以偏概全的行为。但是 NFS 自己有误差却不提,我用我的理想值和你的不好的实际值进行对比,这是什么行为?而更关键的是,即便测试数据符合 Klippel 自身判断的拟合标准,测试结果也有可能存在误差。因为这些人的理由是消声室低频是不准的。当然,最开始的视频中我也介绍了,把低频甚至高频他们说的相位不准的频段也去掉。而为什么中频也会存在偏差?比如我这里举了一个例子,这个音箱在NFS 中测试是这样的结果,但是在消声室中用 Soundcheck 测是另外一个结果。这里比较抱歉。不同的测试仪器导入到相同的一个对比框中,确实比较麻烦。我比较偷懒,大家可以暂停去看它们之间哪些频段有偏差。
更进一步,我这里用的是 Klippel 的另一种型号的音频分析仪。这一支音箱在 Klippel 的另一种测试仪器与在 NFS 测试中,结果也是存在不同的。这两个例子并不是特例,这是我随便找出来的就发现存在偏差的。而且这里既不是低频偏差,我们也不管他们高频的偏差,在中频部分,他们也是存在偏差的。而且如果仔细看,应该可以发现这两支音箱它们偏差的频段也是不一样的。
所以这里我有理由去怀疑这个偏差是NFS 自身测试结果的误差,而不是消声室测量存在误差。因为这既不是低频,也不是高频。而且我可以在不同的消声室用不同的测试仪器测试到相似的结果。而更深层次的原因是因为在销商室中测试是实际的测量值,而 NFS 所导出的频响曲线是计算值,是预测值,不是测量值。方便大家理解,我这里播一段视频片段。
(视频片段来自CCTV新闻频道的面对面节目,丁仲礼:什么是公平的减排方案)
柴静记者:如果它模拟计算出来,这一切是可信,不也是一个依据吗?
丁仲礼院士:你怎么知道它可信?
柴静记者:我们几乎是信仰实验室里所有数据。
丁仲礼院士:它不是实验室,它是计算机,你怎么知道它可信还是不可信?
柴静记者:丁院士,我们当然知道科学界有反对和怀疑的声音,但是给我们的印象是,因为 IPCC 的这样的一个研究的组织,它也是各国的科学家在一起拿出一份报告。而且也是因为有报告做基础,全世界的国家会到去开一个气候的大会,所以给我们的印象他是得到了主流科学界的认同的。
丁仲礼院士:科学家有主流吗?
柴静记者:我们理解的主流是……
丁仲礼院士:科学家是根据人多人少来定的吗?科学是真理的判断。
有些测评媒体会说存在和消声室匹配很好的结果,或者一些 NFS 测试很平直的结果。首先这本身就是一个悖论,因为按照这些媒体的说法,既然消声室是不准的, NFS 如果和消声室测试结果一致,不就恰恰说明 NFS 也不准了吗?而 NFS 的测试结果平直,就一定能说明测试结果是准确的吗?从我刚才的例子中就能看到,如果消声室和 NFS 的计算结果存在偏差,我认为完全有可能存在 NFS 计算出来的结果是平直的。但是消声室中或者现实中不平直的例子,比如我刚才把纽曼的 KH120 的音箱加一个 EQ,让它在 NFS 的结果中看起来平直。即便这样,可能在消声室中存在更大的偏差。
我相信Klippel自己应该是不会说出这么业余的论据的,因为不论是判断测量值还是计算值,都不能简单的用这样的个例作为依据,尤其是存在有偏差有冲突的一些结果。我自己测试的一些专业扩声音箱也有 NFS 计算结果和消声室测试几乎完美匹配的例子。但我不能因为存在这些理想的结果,就忽视掉存在偏差的结果,更不能用一些理想的结果去得出 NFS 的计算结果一定准确的结论,尤其是在没有任何对比的情况下。如果大家还不明白,我这里举一个不恰当的例子。
NFS 这样计算出来的值就像是《三体》游戏中一些 NPC 角色所自认为的一些万年历,而实际的测量出来的一些有偏差。它的情况相当于《三体》游戏中的飞星现象,而且实际上概率远比三体世界中的飞星的概率要高得多。我不能因为我要说明万年历是没有偏差的,就有选择的,忽视掉这些飞星现象。也不能因为这本万年历在一定时间内确实预测到了三体运动的规律,就去认为之后的三体运动也一定会遵循这本万年历。
我不能因为预测了一部分正确的结果,就冒充伟大的先知。况且在机械臂的内容中,我就已经说了,输入的原始数据就存在误差,再加上刚才提到的算法本身可能存在一些问题,负得正,得到相对比较正确的结论也许才是偶然的事件,而错误是必然。回测试仪器准确性的判断和标定。现实世界中的测试仪器并不是圣诞节玩具,你说它准不准,是要遵循基本法的。通常一个测试仪器,它的精度或者准确性是要按照这样的一个体系去进行标定或者是验证的。像 Gras 和 AP 的产品都是经过了严格的这套体系去说明它的精度,而不是我自己说我精度很高,我没有任何的对比或者任何更高 level 的这种标准去进行标定,而不是像 NFS 这样的。在我自己的软件中,我显示我的拟合是有效的,我输出的有效值,所以我这个时候输出的计算出来频向曲线就是没有误差的。比消声室更准的。怎么能这个样子。一个测试仪器,它的精度需要比它精度更高一级这样的标定单位进行标定和验证。而如果你认为你比所有人都要精度更高,你要获得国际度量衡大会的认证,而不是你自己说自己很精确。
而且这里我还看到有一些人会诡辩, NFS 上面用的 Gras 麦克风是校准过的,或者是灵敏度的校准。校准结果不能等同于 NFS 最终计算输出的频响曲线与标准值之间的偏差。这完全是两个概念。刚才所说的是对传感器校准,而这里我们要用到的是 NFS 最终计算出来的频赏曲线的输出结果,是这个结果和其他一些更标准的度量设施之间进行标定和验证。而对于测试仪器的可靠性和一致性也有相关的国际标准。就像我刚才所说, AP 和 Gras 的测试设备有这些完善的,我可以拿得出来证明我确实是符合我所宣称的精度的这样的标准,或者说这样的过程。如果 NFS 也有这样的比较严谨的一个标定过程,我认为这个过程或者流程应该公开出来。而对于一个输出计算值的系统,哪怕像Sean Olive的音箱预测模型,也有置信区间,也有置信度,也有拟合误差。
而且比如这张图,你不能用有的预测值和实测值完美的落在曲线上,就说明这种预测方法本身是没有任何误差的,尤其是存在其他偏离预测方法的点。如果只用拟合曲线上的点,而忽视掉其他有偏差的点,实际上是为了证明这个结论准确。而证明这个结论准确,在底层逻辑上就是错误的。
注意,我这里说的拟合误差并不是 NFS 里面显示的拟合误差,而是 NFS 测试仪器输出的计算值和相对来比较标准的消声室,或者可以用于标定的系统的测量值之之间的对比。 NFS 的计算值与实际的测量值到底有没有偏差?如果有这个偏差大概是多少?这个公差范围是多少?而不是像这样的测试结果中显示出一些不存在的,在消声室中也测量不到的谐振。显然,仅从这里的测试结果也能够看出,已经超出了 Klippel 所宣称的公差范围。
NFS 有些时候就是会测试出一些本身不存在的谐振。这个视频说到这里已经很长了,还有非常非常多的问题,如果以后我有机会再讲。比如有些时候即便在设备前期测试校准的时候是信噪比都没有问题的,但是最后测试完成之后,它自身显示的拟合误差就是不可用的。而这就是在一个正常的房间中去测试的。这也就说明要 Klippel 所宣称的,他们可以通过算法滤除房间的干扰。这样的宣传是不成立的,至少不是每一次都行的。要么就说明明完成了这样的计算,但是拟合误差里显示却有问题。实际上如果只是单独看曲线,尤其是中低频的部分,大家可以看出其实它跟正常的音箱没有什么区别。所以如果拟合误差本身显示都是有问题的,我实际上如果在没有对比的情况下,根本就不知道这个时候它计算出来值到底是对的还是错的。
也许有些人也会说,你讲这么又臭又长的视频,为什么不直接跟 Klippel 去讲?你在这里公开讲这个的事情,是不是要针对他怎么样的?其实我所说的所有的问题,包括更多的问题,我之前都跟 Klippel 反映过不止一次。比如两台不同的NFS,他们测同一个样品,低频就有 2 DB 的偏差。但是这个样品在三个不同的消声室中测试,基本上都是重合的。产品本身也不是我一个人经历,只不过因为是之前公司研发时的项目,不方便直接公开在这个场合讲。但是我这些问题都跟 Klippel 反映过,甚至在 2019 年的时候,我就找过他们的售后,但是很多问题都是不了了之。再比如我这一期视频中所讲的很多问题,也不只是我一个人发现,我做视频也不是为了抬杠。
所以在视频的最后,我给出几个我认为有建设性的意见。首先第一点, NFS 在测试的时候需要和比较好的消声室进行对比和验证,如果结果比较相似比较好。如果结果是不同的,需要找出到底是哪里导致的不同,到底谁准谁不准。第二个点,如果他们的结果是相似的,可以粗略或者大体上的 NFS 能够达到消声室的效果,这样简单的来说是没问题的。但是如果你一定要较真说 NFS 一定比消声室更准,这个结论是不成立的。第三个建议是,我认为Klippel要么改进自己的机械臂,要么改善完善自己的算法,把这些问题解决,让这些 bug 变得更少,而不是回避或者一味的宣称我这个东西都是没问题的。其实讲这样一期视频,除了又臭又长以外,也得罪人。为什么之前发现这些问题的人都没有像我这样公开去讲呢?毕竟谁也不想得罪这样一位德高望重的老教授。而且要讨论这个问题,我建议最好还是从声学原理和测量的原理去讨论,而不是哪个人说了什么。因为如果单纯用某个权威的人说了没问题,直接说是因为 Klippel 发明了,所以肯定没问题,还讨论这些技术干嘛?
Attachments
Last edited: