Regression with v0.2: Filtering of keyboard strokes no longer working properly

With the new v0.2 I noticed a regression in the filtering of keyboard strokes from mechanical keyboards. rnnoise now fails to filter agressive, loud typing noises. I've since downgraded to 1cbdbcf1, which fixes the issue.

Any idea what could cause this?

It could be about the training data, which was recently changed to exclude all non-free data. Could you give some other versions a try, e.g. e10a669e 8e5d4806 59178707 and see if they work better? In all cases, make sure you re-run autogen.sh to ensure the corresponding model gets downloaded.

Couldn't test them live - I'm using rnnoise with easyeffects, and it fails with undefined reference to 'rnnoise_model_free'.

Instead, I recorded a sample file (https://fb.hash.works/zYRJrEMk/) and tested it with rnnoise_demo. Besides speech, it contains two tests on a mechanical keyboard (MX Brown keys):

• Test 1: Normal typing
• Test 2: Aggressively hitting a single key

I converted the files to and from PCM using ffmpeg:

• ffmpeg -i test.wav -f s16le -acodec pcm_s16le -ar 48000 -ac 1 test.pcm
• ffmpeg -f s16le -ar 48000 -ac 1 -i test_1cbdbcf1.pcm test_1cbdbcf1.wav

Commit	Model Version	Speech	Test 1 Result	Test 2 Result	Link
1cbdbcf1	Unknown	Normal	Barely hearable	Barely hearable	https://fb.hash.works/U08kI0jw/
e10a669e	cca6607d	Normal	Barely hearable	Hearable, but fades out quickly	https://fb.hash.works/8Jz5D5BD/
8e5d4806	88a90754	Normal	Barely hearable, echos	First stroke hearable, rest filtered	https://fb.hash.works/ax7dZWt/
59178707	f56003f2	Normal	Filtered	First stroke barely hearable, rest filtered	https://fb.hash.works/CPMih/
904a876d (v0.2)	0b50c454	Normal	Barely hearable, echos	Hearable, but fades out slowly	https://fb.hash.works/DT2eZuT/
2e3c812c (master)	0b50c454	Normal	Barely hearable, echos	Hearable, but fades out slowly	https://fb.hash.works/raGKK7Wx/

Full archive with all WAV and PCM files: https://fb.hash.works/m-B6OXceEG/tar

As you can see some commits you mentioned are a bit better than v0.2. In test 1, 59178707 even wins over the old 1cbdbcf1. But in test 2 the old version still excels.

Just pushed some augmentation improvements as well as new models. That seems to fix the test clip you have and improve things overall. Let me know how things go. If you have any other useful test clips, I'm also interested.

I've tested the exp_fgnoise2 branch for a week now, it seems to fix the problems I had. Thanks! Feel free to close the issue once you merge and draft a new release.

It's now merged in main, along with some more recent changes.