Thank you for sharing this so interesting information, I passed a good time reading Olive et al. article.I'm specifically talking about computed preference scores based on research by Sean Olive and Todd Welti (so is @Deluxillo, I think).

I had to search many audio related terms which was very formative, some of them were confusing when reading at first time.

My only criticism to Olive’s model is the statistical analysis and the whole model based on linear regression. The problem of which maybe in the core of many debates here in ASR.

Linear regression methods tend to find a minimum (optimal metric approach) distance to a hypothetical idealized function that describes a real system.

The problem of this methods regarding to human taste is the non-unicity of the solution in real world, which demands to non-linear models that give more than one solutions.

I explain this more simply by the now infamous “tomato sauce theorem”

Some decades ago, a company producing a popular ketchup sauce, engaged a team of researchers conducting a study of which variables could predict the very best and most accepted ketchup recipe.

After making independent variables as acidity (pH), sweetness (glucose concentration) and others, they conclude one linear model, as Olive did, based on a first round taste of lots of different sauces and punctuation about variables perceived by the subjects, and the ulterior regression model to weight the coefficient.

The real marketing of the predicted ketchup was very poor with respect to expectancy of the brand: less then 6/10 satisfaction.

The experts thought deeply and made new studies with the same volunteers, and came to a surprising (not today but at the epoch) conclusion: they were no one single solution, but a series of clusters of people in which different models were highly satisfying among subgroups: some models weighed more on acidity variable, other on texture and so forth…

As a conclusion, 3 or 4 different sauces was revealed to give a global 9 to 10/10 satisfaction and no single model scored beyond 6/10, revealing the non-linear behavior of human preferences.

Possibly this applies to Olive model and other Harman curves, as I can saw they are all based in the ancient linear supposition of predictability. Olive himself noticed that some underlying dependence between variables could lead to model limitation, but I was surprised that he wrote linear regression as a strong trustable statistical analysis (which it is, applied to many systems), whereas barely all industries today produce various subtypes of same product (see soda companies for example) in order to match cluster-like preferences of consumers.

But still a good and formative article, ASR is providing me a lot of new knowledge, most of which is far from my economic possibilities to test

POST EDITING: I now reading about a family of curves that tend to match different tastes, classifying the listeners in subgroups: sorry, I begun with ancient articles. My critique is only applied to the original Olive et all model and early Harman curve.

Last edited: