roughness_dw time performance to be optimized?

[Antoine101]

Hi folks,

First of all, thank you for providing access to such a great library!

I am trying to implement the calculation of psychoacoustics metrics in my sound classification machine learning project.
I want to calculate a few metrics on short snapshots of sound and as I have many sound events, it leads to a huge number of snapshots.

I noticed the calculation of roughness is significantly greater than other metrics such as loudness or sharpness. I don't know a lot about psychoacoustics or how the calculations are done.
Calculating roughness on a single short 200ms signal (=1 frame) takes around 0.5s.
Calculating roughness directly from the spectrum of this single short 200ms signal takes around 2.1s.

So my questions are:

• Has the roughness_dw calculation script been optimized to its full potential?
• What can explain the difference in time it takes between calculating from the time signal or the spectrum directly? I would expect the former to take less time, not more.

Many thanks

Cheers

Antoine

[wantysal]

Hi Antoine,

Thanks for raising this point. There is indeed a performance problem with the roughness_dw calculation.
We'll run some tests and try to figure out where the difference between the 2 functions comes from.

If you have any idea about how to improve the code, don't hesitate to share ;)

Thanks,
Salomé

[Antoine101]

Thank you for reviewing this issue Salomé!

Please let me know when and if you come up with an optimisation of the code.
I will have a look at it on my side as well.

Cheers

Antoine