A speaker's response can alter significantly with environmental changes.  To accommodate this NWin requires the user to supply a physical reference to calibrate the system and then judges test units by their deviation from this reference.  This has the advantage that the system can be periodically re-calibrated by re-measuring the reference accommodating any change in the environment.  (As a physical reference is so important to NWin it can identify production units that closely resemble the reference making it easier to generate backup and duplicate references.)

The next trick is to analyse the deviation in some way to get a meaningful pass or fail decision.

Your nearly perfect speaker has an unfortunate but acceptable 2 dB mid-range blip.

In production this blip has move slightly. You decide this is not important as it's only a slightly shifted resonance - but note that at the marked point the difference between reference and the test is almost the height of the blip.

To ensure the production unit passes you could apply a simple ± 2 dB limit.

Unfortunately, a simple ± 2 dB limit will also pass this unit. The hatched area shows there is major sensitivity difference between this unit and the reference, and you would not want to pass it.

You could try customising the limits to deal with this particular problem.

Limits customised to deal with a particular production characteristic have the problem that they may not deal reasonably with new variations, like this example where the resonance has moved down in frequency and has a slight sensitivity increase at the top end.

You could modify the limit (or mask) to accommodate the resonance down shift but you are setting up a system where you will be continually chasing production variations.  It is also sometimes difficult to envisage the consequences of change.  For example with this mask you could have this unit pass!

To solve this dilemma NWin employs averaging. The check range is divided into regions containing an equal number of points.

The points in each region are averaged giving a plot like this.  It looks like a third octave response except that the width of each region is not limited to an octave or a third octave.  The worst case difference between averages is 1 dB as shown.

If you now apply the same process to the second test unit you get a completely different result and a way of differentiating between the two units.  A pass or failure limit of 1.5 dB would pass the first unit and fail the second.  (NWin doesn't quite do this.  NWin averages a rolling window.  For example if we were averaging regions of 8 points NWin would first average points 1 to 8 inclusive, then 2 to 9 inclusive and so on.  A rolling window is difficult to visualise - we would need a three dimensional display - and as it works the same way as our examples we'll stick to the way we've done it so far!)

If the number of points averaged is increased, the differentiation between the units can be increased even further.  In this example it is doubled.

By averaging you not only determine the magnitude of a deviation from standard but its "width" as well.  The engineer is given a second parameter to help differentiate between bad and good units.  In addition to a dB limit the engineer has the "Block Point" ie the number of points averaged.  Engineers in fact have a third tool. They can choose over what part of a response to apply the test, ie they define a "check range".

To further increase flexibility NaTKiT allows up to six sets of fail criteria, ie LAYERS.  For each layer, the engineer can set the averaging number, ie the "Block Point", a check range and a failure threshold in dB.