1.20 and later regress audio quality and behaviour

[NoTuxNoBux]

Since a few weeks ago Arch Linux updated to 1.20 and later to 1.21. These versions seem to noticeably regress behaviour compared to 1.10 for me:

• Cut-outs occur often where segments of audio in one sentence get cut out.
• Voice sounds noticeably more 'muffled' (a bit as if it is lacking treble) to other participants.

This happens with the same parameters as configured for 1.10, but even lowering the thresholds doesn't work. I'm experiencing this with WebRTC in Firefox using call systems such as Microsoft Teams, but also with Dino (XMPP), for example.

I also did some testing with a WebRTC tester, but for some reason it's less bad there than it is for others. Currently the situation is bad enough for me that I had to disable use altogether. Downgrading to 1.10 fixes it, though, so that's the workaround I'm using now.

My microphone setup isn't special by any means. I'm using stock PipeWire and the configuration from the README with a mono built-in laptop microphone.

[werman]

I found a good recording I've made long ago that has multiple voices with a wide range of volume and noise sources. And, you are right, it seems that new weights I pulled from xiph's repo are much more aggressive in suppressing noise and classify as noise wider range of sounds, e.g. laugh.

Don't want to do that, but I will likely add a selection of models, with old one as default, because it gives more reasonable results.

When I was testing new model, I tested a different class of noises where it didn't regressed and was even better, haven't thought to test it further since I expected it to be mostly an improvement on the old one.