hi all,

I did a software game to train for frequency response problems :
a random EQ is calculated and the listener has to find the inverse EQ.
here it this soft : resone
A score is calculated to quantify how good the correction EQ has been done.
But I'm not allways satisfied about the score value compared to the "psychoacoustic" result :
- sometimes a high score sounds far from perfect
- sometimes quite a good correction EQ gets a lower score

My score is calculated as a "surface" between corrected curve and flat curve, weighted with frequency : a sort of (amplitude x bandwidth / center frequency).
Maybe what I should do is :
- a "weighting" of the score depending on center frequency of the irregularity, midrange irregularities are more audible than low freq or high freq problems
- as higher amplitude with less bandwidth is less objectionnable than lower amplitude with greater bandwidth irregularities, add a kind of weighting depending on the (amplitude / bandwidth) of the problem

I'm not sure that my explanations are so clear, but I'd like to know if someone has got reference articles about quantifying audible frequency irregularities.

Thanks for any info.

You would need for this something like implementation of ITU-R BS.1387 (PEAQ). There are two models: Basic and Advanced. There are publicly available implementations of Basic model (EAQUAL and PQevalAudio) and Opera is implementation of Advanced model (but is way to expensive).
There are also other method for objective evaluation of audio quality, but ITU-R BS.1387 seems to produce best results. Yet in my opinion PEAQ is not accurate enough and thus probably will not be good enough for your needs (it was trained on different kind of distortions than what you are trying to measure and is accurate enough only on samples similar to training set).

thanks muaddib for your answer,
I think you're right in saying that PEAQ is implemented to measure mainly other audio problems than amplitude/frequency variations. So in my case, I don't think it is the best solution.
Any other info ?
Bye

It's not that simple. How do you define "flat"?

Is it "flat over long averaging period" or "flat first attack" or what?

You can correct one or the other, but not both, unless the combination of your speaker and room have the same amplitude and power response.

The purpose of the soft is to check audibility of amplitude anomalies (peaks or notches), so it is generally used with headphones, so no room response.
In the case of a louspeaker in a room, it's generally more difficult to detect the added random EQ but it doesn't change the way to test : you can allways compare the input signal with no EQ to the signal with the random EQ.
If you try the soft,, my question is easier to understand. For example, a peak of x dB with a given Q is generally more audible than a notch with same characteristics. But what kind of "weighting" should be used to quantify this notch compared to the peak ?

Thanks.

I am going to venture a general comment. It seems to me that you are trying to put a metric on a function space not endowed with a natural one. You are unsatisfied with the ones you have written down, because they don't agree with results of your listening tests.

So it seems to me that you need to turn your question around and think of the metric itself as the object you are measuring or attempting to infer. I think you should start with a list of candidate metrics, which each have a number of undetermined parameters, use the listening tests to tune the parameters in each metric, and finally select the metric that best fits the data.

It strikes me that this is a problem of Bayesian inference, a subject I know little about. You might check out that literature for help on how to make the advice from the previous paragraph more concrete.

There is also a hidden assumption in the original idea that people will agree on what is best, in any flawed system.

I think that it's clear that for any non-transparent system, different people will either prefer or dislike different things.

In my case, the non-transparency is only one peak or notch in the frequency response. And I don't need a very precise weighting of the "uneveness".
I'll try the suggestion from mattc find and tune metrics that correlates best with psychoacoustic results.
But I thought that maybe somebody already did this kind of job