Still, the real test of Audiolense for me will be in my living room setup - unfortunately I haven't yet gotten around to doing that.
So I've finally managed to do some more testing - this time specifically focusing on Audiolense in my living room setup.
Have to say that it did take a little bit of reading and a few tries for me to get consistent measurement results with Audiolense, but I believe I got there in the end
To try and make my results more meaningful I did some comparisons too, as well as some blind listening tests with foobar2000 ABX utility.
Audiolense XO - living room
After using the SW a bit I'll rate usability as just 'OK' - it is definitively not the most user friendly layout and UI, and I believe most users will need a while to find their way around it. However it seems there's a lot of power and customizability built-in, so I can understand the appeal for power users.
Anyway, after a few takes I got a pretty clean measurement (single-point), performed noise filtering, used the default true time domain correction algorithm and the Olive/Toole target curve/tilt this time around (i.e. slope of -1 dB per octave or -10dB total slope from 20Hz to 20kHz). This is what the SW simulated result looks like after filter calculation compared to measured response:
Using REW to validate the correction resulted in basically identical measured response!
I could hardly believe my eyes seeing alignment this good:
Similarly for the time domain, simulated step-response after correction:
Vs measured step response:
Practically identical. So I guess we can very well trust what the SW predicts
Let's see if the step response degrades if we move out of the MLP horizontally:
Seems tweeter and woofer remain aligned for the most part, but we do get some pre-ringing if we move ~70cm to the side of MLP.
For comparison, here's how this looks with Dirac Live:
As we can see there's a bit less time-domain correction done by Dirac, but there's also less variation as we move away from MLP.
Lastly here's the speaker's uncorrected step response for reference - clearly showing unaligned tweeter and woofer responses:
Audiolense here adds almost 0,4s delay to the signal to allow for non-causal filters and therefore phase corrections. This can be seen from the impulse response simulation window:
For comparison, Dirac live seems to add much less delay to the signal - around 8ms (image reposted from
post #32):
Here's how the generated filter for left channel looks:
The phase correction is very visible indeed. Phase slope is steeper than what Dirac came up with previously.
Speaking of that, let's see how similar are the Audiolense filters to those of Dirac with the same target (left channel only):
Close in general, but Audiolense decides to fill-in more dips - we can see that from having more peaks in the filter response (e.g. 85-100Hz, 200-250Hz, 600Hz, 950Hz) - more on this next.
What was really bugging me was how well would this correction translate to other measurement points across the couch.
With other DRCs I definitely saw issues at some measurement points away from MLP, if the correction was based on a single-point measurement at MLP only.
So to check I tested by doing 6-point sweep measurements across the whole couch (to show wide listening area behavior) and also MMM periodic-pink noise measurements around the MLP (to show the spatial average close to the MLP).
Left speaker multi-point sweeps (wide listening area response):
Right speaker multi-point sweeps (wide listening area response):
MMM measurements (MLP-focused listening area response) - speakers without correction (dotted, upper lines), and with Audiolense (solid, lower lines):
As you can see, although still an improvement, these sadly do not look as nice as the MLP single-point measurement made it look - some peaks show through, and these of course correlate well with peaks in the filter response measured a few figures above. I guess this is just the nature of single-point measurement - the SW probably cannot determine from just a single measurement which dip is safe to correct and which not.
Listening impressions
Initial listening impressions with Audiolense were pretty positive, definitively a clear improvement vs without any room correction. However, as I listened on, and depending on tracks I listened to, I noticed some resonances and 'blurring' of the bass response here-and-there, which, now looking at measurement results, seem could be related to the peaks at ~110Hz and ~170-220Hz visible in the spatial averages. Still, it was not major, and it was quite enjoyable to listen to overall.
Anyway, more on all this in the comparative listening test at the end
Comparisons with Dirac Live and MathAudio Room EQ
Let's now compare measurements done with corrections based on equivalent target curves with Dirac Live and MathAudio Room EQ (i.e. Olive/Toole target). In both Dirac and MathAudio cases multi-point measurements were used to calculate corrections, unlike Audiolense where I could only use a single-point measurement (since there's no way to save measurements and use them in a multi-point pool in the Audiolense trial).
Let's start with MLP single-point sweep measurements,
Dirac Live:
Improvement definitely visible vs no correction but it is a much more jagged response than Audiolense at same location (MLP).
MathAudio:
Similar to Dirac, perhaps a bit more even - still not close to AudioLense.
Now let's see how spatial average around the MLP looks, measured with MMM:
Dirac Live:
Notice no peaks sticking much out of the curve, unlike with Audiolense. Guess this is where some benefits of multi-point correction come through
MathAudio:
Same deal as Dirac - spatial average looks pretty smooth, maybe even more so!
Here's Dirac Live correction measured at 6 spatial positions across the couch with sweep measurements, and their averages:
Left speaker:
Right speaker:
NOTE: Please disregard the HF downward slope >10kHz - I guess I didn't load the 90° mic calibration curve in REW for these measurements.
Let's just compare the average curves between Dirac Live and Audiolense (right speaker shown in upper curves, left speaker shown in lower curves):
I guess we can pretty much see the pattern now - with Dirac there's less peaks going much above the target response curve on individual measured or average curves compared to Audiolense.
Again, I attribute this to multi-point vs single-point correction approach rather than being some limitation of the SW itself - however since I can't test Audiolense multi-point correction I can't verify how different Audiolense multi-point correction would look.
Blind listening results and impressions
First let me start by saying that testing all these various DRCs, my main takeaway was that IMHO they can all be made to sound more similar than different with the correctly configured and equivalent target curves. Any audible differences were in most cases subtle, and naturally I was very much skeptical whether these differences were even real, or just a figment of my imagination
So I decided to do a blind test to see if I can differentiate between various DRCs with any consistency. To perform the test I did the following:
- Exported a track sample file (90s) without any correction, and then exported equivalent sample files but now each processed with one of Audiolense, Dirac Live or MathAudio Room EQ corrections
- Imported the tracks into PreSonus Studio One project view, to match loudness using the LUFS scale within +/-0,1 LU
- Loaded the tracks into foobar2000, and used Replay Gain to double-check and fine tune loudness matching
- Disabled any global EQ or room correction, so only the correction now embedded in exported files is in use
- Used foobar2000 ABX comparator component utility to compare the files. I used 12 trials for each run.
- The files were of course played and ABX compared over my living-room stereo system for which the corrections were generated.
NOTE: It could be argued that loudness matching of spectrally different versions of the same file (i.e. EQed differently, as here) is a much more difficult exercise than just using average loudness scale such as total RMS dBFS or INT LUFS as I did. Though that may be true, my reasoning was that there's still a lot of correlation and similarity between various corrections, and that overall loudness will still perceptually be fairly closely matched between files. Subjectively, my listening impressions confirm this - loudness perception was the same between matched files for me, with only slight tonality differences between various DRCs - i.e. as expected.
First two runs were supposed to be pretty simple, just to see if I can consistently pick out the corrected vs uncorrected files.
- 1st run: Track: Josephine Cronholm - In Your Wild Garden, comparison of Audiolense corrected file vs no correction applied.
Code:
foo_abx 2.0.6d report
foobar2000 v1.6.2
2021-01-11 20:06:05
File A: Audiolense - Josephine Cronholm - In Your Wild Garden.flac
SHA1: cd2c43446eee7b60e59b5e1458dd206bbca511b6
Gain adjustment: +6.95 dB
File B: Josephine Cronholm - In Your Wild Garden.flac
SHA1: a7a32edc7bd678368d57ce7136564986b50fe369
Gain adjustment: +6.90 dB
Output:
Default : Primary Sound Driver
Crossfading: NO
20:06:05 : Test started.
20:08:20 : 01/01
20:08:33 : 02/02
20:08:44 : 03/03
20:08:57 : 04/04
20:09:10 : 05/05
20:09:23 : 06/06
20:09:34 : 07/07
20:09:46 : 08/08
20:09:59 : 09/09
20:10:11 : 10/10
20:10:24 : 11/11
20:10:35 : 12/12
20:10:35 : Test finished.
----------
Total: 12/12
p-value: 0.0002 (0.02%)
-- signature --
f3a66f45a8b842f950ee5f18d147a050c7596307
With this track I get an awful lot of bass resonance without correction, so it was pretty easy to pick out. Phew, that was a relief
- 2nd run: Track: Toto - I WIll Remember, comparison of Dirac Live corrected file vs no correction applied.
Code:
foo_abx 2.0.6d reportfoobar2000 v1.6.2
2021-01-11 19:44:53
File A: Dirac Live - Toto - I WIll Remember.flac
SHA1: 943231d65994cd11bb2048aee8bb941007956f20
Gain adjustment: +3.48 dB
File B: Toto - I WIll Remember.flac
SHA1: c332fec9d8a67a403d5a0186cd6874d41a9082fa
Gain adjustment: +3.44 dB
Output:
Default : Primary Sound Driver
Crossfading: NO
19:44:53 : Test started.
19:49:33 : 01/01
19:51:33 : 02/02
19:54:41 : 03/03
19:55:20 : 04/04
19:55:38 : 05/05
19:57:11 : 06/06
19:58:26 : 06/07
19:58:51 : 07/08
20:00:15 : 08/09
20:00:48 : 09/10
20:01:18 : 10/11
20:01:37 : 11/12
20:01:37 : Test finished.
----------
Total: 11/12
p-value: 0.0032 (0.32%)
-- signature --
2ab2a170135d1318e5036beb67a81d772166323f
This track also energizes room resonances for me, but to a somewhat lesser degree - while the difference is audible, it was easier to get confused this time around as one gets tired as the test progresses. This happened to me already after 5-6 trials, and I had to try hard to focus and continue. Definitely harder than last one to pick out
- 3rd run: Track: Allen-Lande - Gone Too Far, comparison of Dirac Live and Audiolense corrected files.
Code:
foo_abx 2.0.6d reportfoobar2000 v1.6.2
2021-01-11 20:14:06
File A: Dirac Live - Allen-Lande - Gone Too Far.flac
SHA1: c2fa5f34778533443009bffe9c424d77d0d4e605
Gain adjustment: -0.96 dB
File B: Audiolense - Allen-Lande - Gone Too Far.flac
SHA1: 115fdb9115cb22a1bcb73b071f5e23ab493cf709
Gain adjustment: -0.85 dB
Output:
Default : Primary Sound Driver
Crossfading: NO
20:14:06 : Test started.
20:16:01 : 01/01
20:16:12 : 02/02
20:16:28 : 03/03
20:16:42 : 04/04
20:16:57 : 05/05
20:17:05 : 06/06
20:17:13 : 07/07
20:17:20 : 08/08
20:17:27 : 09/09
20:17:32 : 10/10
20:17:55 : 11/11
20:18:10 : 12/12
20:18:10 : Test finished.
----------
Total: 12/12
p-value: 0.0002 (0.02%)
-- signature --
642dfa7bb365cbf5ce3743db32dce3e1558800aa
Used this track as it is a pretty busy, compressed mix, which typically makes it easier for me to pick-out smaller differences in tonality (closer to pink-noise, I guess ). With casual listening I thought Audiolense correction had more energy in the mid and upper bass (almost resonances at times) vs Dirac Live, and this is what I tried to use to anchor myself when doing the ABX. Phew, again
- 4th run: Track: Allen-Lande - Gone Too Far, comparison of Dirac Live and MathAudio corrected files.
Code:
foo_abx 2.0.6d reportfoobar2000 v1.6.2
2021-01-11 20:20:24
File A: Dirac Live - Allen-Lande - Gone Too Far.flac
SHA1: c2fa5f34778533443009bffe9c424d77d0d4e605
Gain adjustment: -0.96 dB
File B: MathAudio - Allen-Lande - Gone Too Far.flac
SHA1: 59409b85ae440e9856e32e5c01fcd16e8a3c160f
Gain adjustment: -0.97 dB
Output:
Default : Primary Sound Driver
Crossfading: NO
20:20:24 : Test started.
20:22:39 : 01/01
20:22:46 : 02/02
20:22:51 : 02/03
20:22:57 : 02/04
20:23:06 : 03/05
20:23:14 : 04/06
20:23:22 : 05/07
20:23:31 : 06/08
20:23:35 : 07/09
20:25:26 : 08/10
20:25:35 : 09/11
20:26:10 : 10/12
20:26:10 : Test finished.
----------
Total: 10/12
p-value: 0.0193 (1.93%)
-- signature --
15253493a16f7500c08ce1c95776b7818d827f60
Used the same track again, for the same reasons. In my previous informal tests (and in above measurements) Dirac and MathAudio corrections seemed more similar, and this was confirmed for me by this ABX. What I used to anchor my picks was a difference in upper-bass response, where I thought I heard a bit more energy in what turned out to be the MathAudio corrected file. It was definitely close and more difficult to pick out than the rest, resulting in 2 mistakes out of 12 and p-value of almost 2%. I guess I'd have to do more trials here to have more certainty in my ability to pick them out
As tedious and tiring exercise this was, it did give me some confidence that the differences I thought I heard with sighted listening were real.
To be honest, considering how I thought the differences between DRCs were pretty small when sighted, I was half-expecting to fail miserably in a blind ABX, so this result was a bit of a positive surprise for me
Another thing I believe is worth noting, is that I could only differentiate different clips based on spectral/tonality differences - i.e. I felt I could hear broad differences and resonances in frequency vs amplitude responses, but I heard no differences that I could correlate to any phase or timing correction differences. E.g. I felt Dirac Live and MathAudio sounded close and they had very similar measured frequency responses, although Dirac does time-domain correction and MathAudio doesn't.
Perhaps I just don't know what to listen for, perhaps the differences would be more obvious with multi-way drivers with digital XOs, multi-channel systems or systems with subwoofers, or maybe these time-domain response differences are simply not very audible in general (as some research on the subject seems to suggest,
e.g. some info in this topic). Personally I won't argue for or against any explanation as I have too little experience on the subject, but will say that this experience does tell me to personally not worry too much about time-domain correction in my own stereo setup
Closing thoughts
Have to say I feel I've learned some more with each iteration of these comparisons, and this one was no exception
In general, my key takeaway is that with appropriate configuration and equivalent target curves, one can get what I feel are pretty similar audible results in most setups with either Audiolense, Dirac or MathAudio.
Most of the audible differences I heard in the end between Audiolense vs Dirac and MathAudio were IMHO related to differences intrinsic to multi-point vs single-point correction approach, rather than having to do with major differences in each SW implementation. Sadly I could not test Audiolense multi-point correction to verify
, but based on all the other results I posted so far, I feel it likely my conclusion wouldn't change much.
In my experience through all these tests, multi-point corrections gave me a more consistent result when listening in the far-field (though it is a compromise at every listening point), whereas with single-point measurements I could sometimes hear some resonances - depending on where exactly I sat or which track I listened to. Maybe this could be further optimized somewhat in Audiolense even with a single-point measurement, but I doubt that in most acoustic environments one can get ruler-flat frequency response at a single measurement point without sacrificing the response somewhat at other points around it.
I'd also like to repeat that all of the audible differences I did hear were spectrum/tonality-related, and that I can't say that I heard any conclusive differences related to phase/time-domain correction done by Audiolense or Dirac in any of my tests - in any case none that I could then anchor myself to and recognize in blind listening. Perhaps others could do better with this - my ears may be far from golden
That being said, the bulk of the very real differences between Audiolense, Dirac, MathAudio and others for me is mainly in their ease of use, measurement/configuration/correction process robustness, configurability/customizability and price. Still, I'd say these are anyway very important considerations in themselves.
Audiolense XO seems to provide much more configuration options than the others, making me think it could be suitable for a power or professional user to calculate custom-made corrections in cases where other automatic systems might give sub-optimal results - and do that on top of also doing XOs and complex driver time alignment while providing the most freedom with regard to target curve tweaking (of the DRCs I tested). However, it is IMHO also more difficult to use and understand than others, so might be overkill for many users and simple setups. The UI is unappealing and dated, and note that you need an additional filter processing engine to apply generated corrections.
On the other hand with
Dirac Live you have a relatively simple, nicely guided and illustrated process for measurement and correction, that allows simple target curve editing and which IMHO gives very good results quickly. IME results I got with Dirac were least sensitive to user-error and therefore quite consistent. However, since it does not appear to be possible to configure much other than choose a listening setup and tweak target curves, if you don't like the out-of-the-box results it could be limiting to some users. The Dirac Live processor component is available in a variety of types, including VST plugins, third-party HW devices and apparently a Windows systemwide application is also planned.
MathAudio Room EQ is fairly simple to use as well, but the measurement process is less guided and it is the least configurable of the three. It is however also the cheapest, and even offers a free foobar2000 plugin variant. VST plugin version is also available, for use in a third-party VST host. IMHO the foobar2000 plugin is a very good option to get familiar with automatic room EQ for free (and probably good enough for many PC stereo users).
There.
I believe with this installment I'm ready to finally put an end to my automatic room correction/EQ investigation saga