There are two broad schools of loudspeaker design: 'rationalists' who are rooted in measurement* and objectivity and the subjectivists who design purely by listening. Most of us fall somewhere in between. We wish that measurements could tell the whole story (if they could, we could design on a laptop on a tropical beach, azure blue sea, endless G&Ts confident that when we sent the design files to Production the speaker would be an immediate success) but they don't. The subjectivists would argue that not only do they not tell the full story they completely fail to describe the listening experience and are, therefore, valueless.
Surely the truth must lie between these extremes. Measurements (frequency response, distortion, dispersion etc.) definitely illuminate some aspect of the design. That is, modern audio analysis computer systems can produce much pretty, multi-coloured even 3D graphical data about the speaker under test. Can we - should we, dare we - trust our eyes when examining these plots? Has an information overload given the designer the illusion that he is more in command of the design than he really is? How is it possible that the BBC was able to produce great loudspeakers in the 1960s entirely 'by hand' a decade before the first calculator? I'd say it was down to careful interpretation of the little measured data that was available combined with a willingness to listen and let the listening experience be the final arbiter - not the formulae and rule book. It may even be that the smaller quantity of difficult and expensive to acquire measured data gives a clearer picture of the speaker's behaviour than the vast quantity of press-of-a-button data that is so readily achievable now. Just as anyone with an internet connection can swat-up and become (their own) brain surgeon, loudspeaker measurement systems are affordable by serious hobbyists; it's a wonder that there are not even more speakers on the market. Factor X that takes a design from the flawed to the great cannot nowadays be related to audio measurement equipment - it must be somehow related to the ability to wade through, discard and interpret the vital, core data. And know what to ignore.
As mentioned previously, we recently relocated Harbeth's R&D facility to a beautiful, spacious 16 century barn. The 500 year old oak beams across the high ceiling have given me the reflection-free heigh to allow me to build an all-weather, fixed, loudspeaker measurement set-up. By manipulating the measurements it's possible to make full-range 20Hz to 20kHz acoustic measurements of any speaker in this ordinary room; something that would have (and did) require a large anechoic chamber before. And I've spent over two months fine tuning the set-up to remove (or at least damp down) reflections so that what the microphone picks-up is as much about the speaker as possible and not the speaker + room. After many long days and evenings fiddling around with positions, angles, heights, path lengths, Rockwool, VetBed wadding and even duvet quilts I have measurements which are as believable as those I took at the BBC anechoic chamber on the same speaker. I can't say more than that. I can't say that there or even the BBC measurements are 100% accurate because there are no absolutes in acoustics. But they at least agree. And that is all one can say. Perhaps I'll write more about the set-up another time.
Objectivist speaker designers place great weight on their frequency response measurements. These, they believe, fundamentally describe how the speaker will perform, and some would work long and hard to achieve a certain shape to the loudness frequency response, plotted on graph paper from the lowest the the highest frequency. Some would perhaps aim for a response that gradually increases in loudness from the lowest to highest frequencies. Another for a basically flat response. Another for a response with a gentle incline from LF to HF. The variables that they'd be playing with would be the latent acoustic characteristics of the drive units (if they actually made the drivers in-house they could play around with the moving parts to their hearts content to achieve the desired response shape) and the crossover/equaliser aka the network.
Now a complex network allows the designer to 'nip and tuck' the system's acoustic output. If he wants to reduce a peak in the speaker's acoustic output (frequency response) he would open an electrical pathway (a controlled, mild short circuit) between the red and black terminals on the speaker box so that less volts arrive at the drive unit in a certain frequency band that it is mechanically rather energetic in - the result would be to flatten an acoustic peak. Filling-in holes in the frequency response is a much more difficult problem because if you could create sound out of no (or too little) sound then the laws of physics would have been beaten. You'd have to reduce the entire audio band by a certain amount to bring the peaks down to the level of the troughs, and in so doing would discard useful sensitivity and introduce crossover complexity, a low system impedance and much cost and size. So in reality, the speaker designer is concerned (or at least, we are) with reducing the peaks, not filling the holes (if there are any), maintaining a good overall sensitivity and high-ish system impedance.
So, when I look at a frequency response curve, I know that my hands are tied concerning filling in any holes or dips in the frequency response by electrical means in the crossover. Those rightly should be addressed in the design of the drive unit - ' a mechanical problem should best be solved mechanically in the drive not electrically in the network' is a good starting position. I'm interested in the peaks, those frequency bands where for often inexplicable reasons the drive units and/or cabinet have extra efficiency, produce more dBs of output. We may be able to bring down those high-energy bands in the crossover/equaliser network.
Attached is a graph of a mini-monitor measured at the BBC chamber and in my new set-up. The red and light blue are basically the same data of the speaker at the BBc chamber in 2008. The yellow is a measurement of the same speaker here, a few days ago. The vertical axis is dB, only 2dB/division (a high resolution display). You'll see that above about 3kHz (horizontal axis) the latest and old measurements are very similar and could have been made even more so had I spent time exactly aligning the microphone. Up to 1kHz the curves are basically the same but between 1 and 3kHz (the presence region) there is about a 2dB difference for the very same speaker. Since the HF (and lf) are consistent over the three years, what we are witnessing is some ageing process in the speaker where the output in that band has increased in loudness. Or is it a measurement error?
Now, back to the point: had I been developing this speaker now in 2011 and had not measured this speaker in 2008, I would have assumed that the peak at 60dB 2kHz was to be pulled down in my network (to, say, the level at about 1kHz, 58dB) and would have thrown away that 2dB in the crossover by using a combination of a coil, a resistor and a capacitor - cost perhaps GBP 7-10 per speaker. But that would reduce the overall efficiency for what is already a low-efficiency speaker system, possibly eaten into the power excursion capability (limited at LF) and only to find ..... that this peak was unstable and moved about in frequency and amplitude as the seasons pass. Hence the danger of failing to take a long, cool, rational view of any speaker measurements.
* When we talk of 'loudspeaker measurement' we mean placing a microphone somewhere near the loudspeaker under test, injecting a test signal into the speaker and plotting a frequency v. measured parameter (loudness, distortion, dispersion) initially on paper chart, now on the computer screen.