Axiom Home Page Forums Thread Like Summary - Speculating on Axiom’s Olive scores

Anyone care to guess the Olive score of Axiom loudspeakers? As audio enthusiasts, I’d thought it would be fun to speculate. If a little bit of knowledge is a dangerous thing, why not play with fire?

As a background, Dr. Sean Olive is an acoustics researcher at Harman and was the principal discoverer of the “Harman curve”, a headphones frequency response that routinely wins A-B listening contests because they’re perceived as having full bandwidth and sounding neutral.

Audio Science Review (ASR) and Erin’s Audio Corner (EAC) have been using a Klippel near field scanner to measure the family of curves of dozens of loudspeakers, and volunteers have plugged the raw numbers into an algorithm to determine an “Olive score”, from negative numbers up to an 8.5. Scores in the 8’s are the best that’s possible, and you’ve reached your loudspeaker end game as far as coloration-free neutral sound. It doesn’t go up to 10 because experienced panelists are reluctant to give any loudspeaker a 10/10 lest something even better comes along, an artifact of the way NRC did their double blind listening tests.

Going through ASR’s catalog of reviews, awful speakers are below 1.0. For reference, an IKEA/Sonos Symfonisk wall speaker is a 3.2 (well into the high fidelity category), a KEF LS50 is 4.6, and Revel/Genelec/Neumann tend to be in the 5’s and 6’s. The highest score so far is a 7.1 for an active DSP loudspeaker. A subwoofer tends to boost scores by 1 to 3 points; everything sounds better with a subwoofer.

It’s been noted by Sean Olive himself (on an EAC YouTube interview) and by ASR’s forum moderators that the Olive score has a loose confidence window of +/- 1.0. That means a 7.0 loudspeaker is “similarly good” to a 6.0 most of the time in double blind testing, but would statistically be preferred in most listening rounds over a 5.8.

From the sound directivity index and listening window curves that Axiom publishes on their website, my guess is that the floor standers (M60, M80, M100, and LFR equivalents) and the M5HP would score around 6.0 to 6.5. There’s not enough information from the bookshelf graphs to speculate. There’s some forum chatter about the limitations of a 2-way design in generating a good family of curves, so it’s possible they’d score in the high 4’s to high-5’s.

Those handful that score in the high 6’s on ASR are active 3-way loudspeakers with DSP crossovers and a rigid metal enclosure. Speakers scoring in the 6’s tend to share the same smooth sloping directivity index and flat listening window that the Axiom curves imply. Scores in the 5’s consistently have directivity errors (particularly at the crossover region), while those in the 4’s tend to have port/cabinet resonances, multiple directivity errors (which show up as changes in slope), and low Q responses (wide but subtle hills and depressions).

Unfortunately, I’m not able to glean this information from Sound Stage Network’s set of anechoic measurements. They publish the listening window but not the sound directivity. It seems like loudspeakers with a reputation for superior measurements have a flat listening window and in-room response curve, and I don’t see that with the Axiom’s and Bryston’s. So I don’t know if my conclusions are correct. In fact, it seems like loudspeakers measured better 20 years ago than today.

The Olive score has limitations, and should be treated as one tool in evaluating a loudspeaker.
- Bad recordings never sound good on any system.
- We prefer multichannel playback over stereo or mono, so we’re probably more forgiving of multiple speaker listening.
- You can EQ some (not all) speakers for better performance.
- In-wall and wall-mounted loudspeakers are penalized for boundary load compensation.
- You can prefer a lower scoring speaker with superior high SPL dynamics.
- Many ASR forum members have openly regretted using the Olive score to influence their purchases. Perhaps scores inflated buyer expectations or they chose hardware too small for their room.

What’s everyone else’s take? Where did I screw up? Does anyone own one of the measured speakers? I’m hoping Ian and Andrew will chime in on our armchair quarterbacking, preferably before we torch the couch.

The patent is an interesting read as it gives a summary the history of different attempts at quantifying loudspeakers (it assumes an understanding of speaker measurements and terms).

From a Speaker Manufacture's point of view, I think, there is only a down side in publishing this number.
Even if your speaker is at the high-end of the scale (unless you have the highest rating one can get) you take the chance that someone could publish one that's higher.
People being human would not take the +/- 1 as being equal ... it would be argued that a higher number is always the better speaker.

Another question of concern: could a manufacture design a speaker to maximize this number without really producing something superior?
Since I only skimmed the patent with the goal on understanding the figures presented and a bit of the math, I only have the gist of this method, In the end, I focused in on figures 5 and 6. I'm still not sure if my understanding of figure 5 (based on an anechoic model developed) & figure 6 (based on a generalized anechoic model developed) is correct. I'm kind of guessing fig #5 is the sampling used (13 speakers) to develop the algorithm and #6 is the sampling used (36 speakers) to test it's predictive measure. If that's the case, overall the blind listen tests correlated with the predictions but it's definitely far from perfect (sometimes it's a lot further off then +/- 1).

I also think there are better ways of coming up with a predictive measure then the linear regression presented ... Easiest would be to apply some pre-packaged AI algorithms . Although, If the sample size presented here is the full dataset (and not just for patent clarity) there might not be enough to train with (on the other hand given 30 independent variables there is a lot to work with). If it's not enough I might consider a few other methods to replace the linear one above. Either way it would probably improve the correlation and depending take a lot of the human tweaking (weighting) out of the equation.

If this method has merit, I think it would only really be useful for manufactures to predict if their speaker would do well in a listening test.
In which case, I'm not sure how it could be monetized as there's nothing to keep a manufacturer from using it as a internal measure (especially if they didn't advertise the fact).
Maybe someone else can fill in what I'm missing here ...