Potential bug when generating RT60 using v0.7.5

[michaeljconnolly]

Hi,

Running the pyroomacoustics_demo notebook, I've noticed that the RT60 generated using version 0.7.5 is significantly different than the RT60 generated using versions 0.7.3 and 0.7.4.

I tested further with my own scripts, that are similar to the demo notebook (both attached here: scripts.zip). One uses ray tracing (raytracing.py), and the other uses ISM (ism.py).

For raytracing, the RT60 is ~40 ms for v0.7.5, but ~280 ms for both v0.7.3 and v0.7.4.

For ISM, the RT60 is 7 ms for v0.7.5, but 31 ms for both v0.7.3 and v0.7.4.

I have reproduced these results using docker containers, built with the provided Dockerfile (changing pyroomacoustics version only).

I wonder if this is a bug, or if I'm missing something?

Thanks for your help.

[fakufaku]

@michaeljconnolly thanks for the heads up!
Actually, I have updated the code for the RT60 to be more robust.
The new code will ignore outlier samples at the tail of the energy curve to try to get a better fit.
Have you tried to check which results makes more sense wrt to the actual RIR?
If you provide plot=True as an argument to measure_rt60, it should also show a plot that should make it clear if there is a problem or not.

[michaeljconnolly]

You're welcome, and thanks for your response!

I have plotted the RIR for one microphone and the RT60 for 0.7.3, 0.7.4 and 0.7.5. These examples closely follow the demo notebook (ray tracing with 1000 rays, max_order 3).

It looks like, the RIR tails off at ~150ms. It is interesting that for the 0.7.5 RT60 plot, the energy does not fall to -60dB.

Images below:

0.7.3

0.7.4

0.7.5

Thanks again for your feedback.

[fakufaku]

Thanks! It looks like the default value of the energy_thresh parameter is too low. Could you please try to set it to 0.99 or 0.999?

[michaeljconnolly]

No problem. I have added some plots for v0.7.5, setting energy_thresh to 0.99, 0.999 and 1.0. The RT60 lines up with 0.7.3 and 0.7.4, when set to 1.0. With energy_thresh set to 0.99 and 0.999, the energy does not fall to -60dB.

Images below:

energy_thresh 0.99

energy_thresh 0.999

energy_thresh 1.0

[fakufaku]

It looks like the energy_thres parameter is more difficult to use than I anticipated.
For now I will change the default to 1.0 to match the previous behavior of the function.

[michaeljconnolly]

Perfect, thanks for your help with this!