Online WPE can’t handle audio signals that have segments of absolute silent

[Bar-BY]

Hello, I ran into an issue with WPEonline: turns out it can’t handle audio signals that have segments of absolute silent (continuous constant value). The way I found out about this was as follows: While running (WPE online class fashion etc…) on a ~2.5 hr session, I got this warning message:

31104it [00:15, 2050.81it/s]//anaconda3/lib/python3.7/site-packages/nara_wpe/wpe.py:790: RuntimeWarning: invalid value encountered in true_divide self.kalman_gain = nominator / denominator[:, None] 37489it [00:18, 1997.07it/s]

Also, when I listened to the output file I noticed it was fine up to about 4 mins in, and then it went completely silent for the rest of the 2.5 hr file. This experience recurred for both multi- and single-channel files. When I took a look at the signal, I turned out that the input signal has a short silent segment right where the output file went silent. here is an example of the input and output plotted on the same time scale

[boeddeker]

Hello,

applying WPE to a signal that has segment of zeros is a bad idea. It may be executable, but the result may be wrong.

Your signal look unnatural, how did you obtained a signal with a segment of perfect silence? There should be microphone noise.

In #44 I change the code, now the code should be executable and yield finite numbers instead of NaNs. But I am not sure if the result will be reasonable. Can you try that PR?

You can test it, when you install nara_wpe from my fork:

git clone https://github.com/boeddeker/nara_wpe.git
cd nara_wpe
pip install --editable .

[Bar-BY]

Hey, I agree, any real recording would never have segments of zeros, this can only happen when the segments are artificially inserted into the signal in some manner and not as genuine microphone sensor output. However, to my own surprise as well, it turns out there are cases where you do get datasets with audio files that contain such segments of zeros due to device issues (CHIME) so it is useful to have to algorithm robust to such things..(yes, it isn't physically meaningful, but it saves the trouble of fixing the signal before running it through WPE) Thanks!

[boeddeker]

Yes, the zeros in CHiME are difficult. But be careful, that my change should make online WPE only executable. I haven't checked that the implementation handles the zeros correct (i.e. drop that frame in the observation). I did this only for the offline implementation.