Audibility of different crossover types

 

Together with a fellow DIY addict, Hans L, I have conducted listening tests on different crossover types. To be able to hear only the effect of the crossover, excluding all other effects that you have with loudspeaker systems, we used two identical midwoofers from SEAS. The reference sound was taken from the unfiltered single upper midwoofer. The tests and observations are described on this page.

Test setup

Loudspeakers: 1 stereo pair with two wideband midbass units on each side (with EQ flat response between 300 and 8000Hz with gentle slopes on both sides of the passband). The two identical units were filtered as a 2-way speaker, one unit taking the role of tweeter and the other doing the woofer.

Source: CD player, digital coax out to Rotel processor/preamplifier which passed the signal on to a EQ circuit (strictly for flattening the midbass response) which in turn went to a M-audio Delta410 PC sound card. Hans L used LspCad Pro v5's Emulator as the crossover, giving us maximum flexibility during the listening tests. The multi-way signal from the delta 410 was amplified by a Harman Kardon receiver, which sent it to the speakerunits.

In short: CD player digital out - Rotel DAC - EQ circuit - PC stereo in (delta 410) - Lspcad Pro emulator - delta410 multichannel out - HK receiver - speakers.

Filterslopes

We listened to the following filters at 250, 500, 1000, 2000 and 3000 Hz:

We also had two allpass filters (tested on a single speakerunit)

  • 1st order allpass at 1KHz (phase rotation equal to LR2)
  • 1st order allpass at 400Hz and at 4000Hz (phase rotation equal to typical 3-way system with two LR2 filters)

The acoustical responses of the two drivers added like textbook filters, despite the fact that their amplitude and phase response was not entirely flat. The fact that the drivers are identical is the key here, allowing us to use textbook active filters for the listening tests. The drivers were sufficiently flat of their own, so although they are not entirely flat, we can safely assume that we're also talking acoustical slopes here when discussing the different crossover types provided by the active processing circuitry.

The reference (REF) was a single unfiltered speakerunit. We listened to all the filters in stereo, but I think we would have drawn the same conclusions in mono. The test stimuli were pink noise, repeating short sound fragments (XLO/Reference Recordings) and several complete pieces of music. General observation: the differences were much more easily heard with pink noise. Music was significantly more cumbersome. Particularly when the structure of the music was more complex, some of the differences were hard to spot because sudden changes in the program material could coincide with the switching of the filter. We also carried out blind listening tests on eachother on several occasions where differences were subtle or virtually non-existant.

Listening

Before starting this test we were primarily interested in the effects of high Q and high vs. low order filters in terms of ringing, intermodulation distortion and others. I had heard too many times from experienced builders that low order was good, very few xo parts was preferable (imho to the point of sacrificing other aspects) and so on. I wasn't sure how we would be able to test these attributes or better yet draw any meaningful conclusions from it, but we went ahead anyway.

After listening to several filters it seemed apparent that the differences in power response compared to the REF were huge. Other effects (ringing, distortion) I was interested in investigating were overshadowed by these deviations in tonal balance. These other effects would have to wait a little...

Results

Before we started this, there was a fierce discussion on the audibility of transient perfect design at one of the Dutch audio forums. The stated amount of audibility varied wildly, from subconscious irritation (listening fatigue) to different attacks/impulses to complete shifts in timbre. I think that I speak for both if I say that we have done our best to hear a difference between the reference and the single and dual allpass filters. But with the test signals we used, we concluded zero audibility. [[[As a side note... we repeated this test (allpass filters) in the fall of the year 2004 with a couple other avid speaker builders and hifi enthousiasts, one of whom is a commercial speaker designer in the Netherlands. This time we tested on an ESL, which was verified transient perfect: with a little help of some acoustic foam we could measure a rather clean square wave in the listening room. We listened to various kinds of music, but were never able to beat the gamble table in a blind test]]]

On to the crossover filters: here are our findings point by point for the xo at 1 kHz:

  • BW1: deviates, just like most of the filters, from the REF. Rather surprising because 1st order filter slopes should be ideal in many respects. With a piece of music the differences are not as easy to hear. BW1 and BW3 have the same character.
  • LR2: A clear loss of energy around the crossover point was audible.
  • LR4 also gives a clear deviation with regard to the balance of the REF. Also here is less power on and around the xo point. With pink noise very clearly audible, with music also, but the perceived differences are not unpleasant, althought slightly lean.
  • LR8 gives a clear improvement on LR4, probably because the transition is narrower, as a result of which dips (in the power response) are less notable.
  • Bess2: Big surprise. Of all filters the closest sounding to the REF using pink noise.
  • BW3 and BW3-i followed the footsteps of Bess2, very similar to REF.

In general the even and odd order filters form groups with similar character in perceived sound. Odd order filters came nearer to the REF... except Bess2. We found the following explanation for this: textbook Bess2 (which is what we're talking about) has a hump of nearly 1.25 dB at the xo point. This on-axis excess energy compensates for the loss of power off-axis with this even order filter.

Our findings for the xo's at 2 kHz:

The general differences in perceived sound are roughly equal to the 1 kHz xo.

  • BW1, large deviation of REF. The interference patterns (vertical lobing) at higher frequencies lead to an unnatural thin sound when compared to the REF.
  • LR4 was associated with 'woofer-tweeter sound', very 'hifi-like' sound, a bit sterile. Smooth but again unnatural when compared to the fuller and coherent sounding REF.
  • LR8 seemed more like LR4 in this case than when crossing at 1KHz. An improvement, but not as large as the relative improvement between LR4/LR8 at 1 kHz. It seems that we come close to an audible threshold when LR8 is used at 1 kHz.
  • Bess2 didn't do as good as with the 1 kHz xo. The sound was a little less copious / rich compared to REF but closer to the REF than LR4 and LR8.
  • Bw3 added something to the sound and BW3-i took something away round about the xo point.
  • BW3 performs very good again, this time even slightly better than Bess2. Closest to the REF!

At lower frequencies we observed that at 500 Hz there was no clear winner. We heard diffences with all filters but all came close to the REF. Maybe you could prefer the LR2 and LR4 in this case as they only take a little bit power away, as opposed to the uneven order filters which add a little bit audible power around the xo point. At 250 Hz, we could not hear a difference with the REF anymore with any filter. Apparently the wavelength was too large with respect to the distance between the drivers. They could be considered to act as a single driver at this xo frequency.

At 3000 Hz we found similar observations when compared with the findings at 2000 Hz. The differences were even a little bit smaller, maybe because the midwoofers start to beam a bit at this frequency and higher. This causes less interference (vertical lobing) issues.

Conclusions

This might come as a shock, but we think that the overall winner was the 3rd order Butterworth filter, which was terribly popular in the 70's and 80's. I think we have heard why this is the case. Another conclusion is that during listening you rapidly gets used to the character of a certain filter. After switching back to the REF, the results can be rather overwhelming, knocking you straight off your feet at times. Even to a point where the REF start sounding... colored! This process can happen within minutes! It happened to us in several occasions. Smart ppl have said it before: aural memory sucks...

Another note: this test was performed by mortals in a real livingroom, not mics in a anechoic chamber. The vertical lobing and related power response obviously becomes moot when listening on-axis in a room without reflection. Such is however never the case. I also understand more of the love for broadband units, but that does not mean that there are no filters that result in an extremely natural reproduction with multiple speaker units. The differences between the winning filters and REF in our listening tests are audible with pink noise, but quite hard to hear with normal music.

What to do now?

We are still thinking about what further to do. What will this experiment bring when we use an actual tweeter instead of two wideband units? To create a REF in that situation is very difficult. Maybe it would be worthwhile to compare the same filter using different distances between woofer and tweeter, as we have found the vertical lobing to be an important audible issue. Other effects, such as ringing, IM distortion etcetera have in fact already been investigated by other very respected people such as Floyd E. Toole. They have carried out way better and thorough tests of course, we only performed some tests in one particular listening room. But the nice thing of our experiment was that I consider it a new approach, to investigate audibility of lobing. The results do not confirm the sometimes heard statement that uneven order filters cause coloration by the asymmetrical lobing, of which I, by the way, have never seen any scientific proof or any investigation (apart from our little experiment).

That's all for now... Maybe there comes a time when we will do some further investigations. If this happens, I will come back ofcourse and update this page.

To complete this page, below I show the equalised single driver response (at different angles) as well as a first and third order Butterworth on-axis, to give an example of the acoustic slopes and how well they add to obtain the same response when compared to the unfiltered single driver. The blue lines indicate amplitude, the green lines indicate phase.

Single driver unfiltered at 0, 15, 30, 45 and 60 degrees

First order Butterworth individual driver response

First order Butterworth, sum of the two drivers

Third order Butterworth individual driver response

Third order Butterworth, sum of the two drivers