[ANN] Noise-repellent 0.1.2 is out

[lucianodato]

Hi folks! It's out! You can grab it here https://github.com/lucianodato/noise-repellent

Changelog for this release:
- Noise profile state saving (noise profile is saved with the session so no need to learn it again) Big thanks to FrnchFrgg!
- Smooth reduction applying envelopes (release control) and a postfilter.
- Fix whitening parameter (now it works as it should)
- Fix adaptive noise estimation (previously known as auto-learn)
- Using Spectral Gating instead of power subtraction when the noise profile is learned manually (behaves better with transients)
- Fixed a bunch of memory leaks. Thanks x42!
- French translations revised by FrnchFrgg. Thanks again!
- All sliders use units for clearer usage

If you are a packager for any distribution and find something odd and needs to be changed for packing it correctly contact me!
There is a deb package here http://www.stockinger.org/musik.html but I didn't do it so It's at your own risk to install it. Hope it's useful for you!

[English Guy]

Do keep us posted, I am quite interested. I have a number of old songs on cassette four track that I intend to try and 'rescue' into the digital domain, so this could be quite handy.
Thanks
Guy

[lucianodato]

Ok so to keep you informed I have more great news! I got the masking model working and it gives really good results. It sounds really good compared to previous versions. Here is the visual proof in a speech file:
https://drive.google.com/open?id=0B7Ydx ... URyX2hYWWM

[lucianodato]

And that is a really noisy file with lots of reduction applied .

[CrocoDuck]

Ok so to keep you informed I have more great news! I got the masking model working and it gives really good results.

Very interesting work! Can I ask where you are getting the information for the perceptual modelling from?

[lucianodato]

I've been reading lots of papers lately. To get this working I had to decrypt lots of them hahha. In all seriousness this is based on a really old method proposed by Johnston (1988) to model masking thresholds for an audio codec. This model is explained in various books and other works. This is a great resource if you are interested http://mp3-tech.org/. It's from one of the LAME mp3 codec devs so it's good info.

[CrocoDuck]

For Arch users: this is now on the AUR.

[lucianodato]

Many thanks!

[English Guy]

Ok so to keep you informed I have more great news! I got the masking model working and it gives really good results. It sounds really good compared to previous versions. Here is the visual proof in a speech file:
https://drive.google.com/open?id=0B7Ydx ... URyX2hYWWM

Thanks. Just caught up with this. The earlier version compiled fine on Debian oldstable and I used it on an old track with a noisy mic and it worked well.

[ufug]

Hi folks! It's out! You can grab it here https://github.com/lucianodato/noise-repellent

Lucian: Just got done with my first time using noise-repellent on some tracks I thought were unsalvageable, and it works great.

Got rid of amp hiss and single coil guitar pick-up hum with minimal impact on the sound. Thank you so much for your amazing work!

[nils]

Could you provide information how that compares to existing noise reducing functions? Like the one built-in in Audacity for example?

[ufug]

Could you provide information how that compares to existing noise reducing functions? Like the one built-in in Audacity for example?

Hi nilshi. It's been years since I used Audacity or its noise filter, so I have no idea how it compares. I just know noise repellent works well and I'm really glad we have it.

[lucianodato]

Hi everyone! The main advantage is that it works in real time, so you have instant feedback on what your tweaks are doing. Audacity uses an attack and release control, noise repellent uses only a release. Noise repellents postfilter in this version is similar to frequency smoothing control in audacity but it's adaptive, based on signal snr so when there is only noise in the signal it will smooth the filter quite a bit more. This means that avoids musical noise between phrases or notes. This is soundwise it's main distinction.
However this is outdated. Now I have working really well a filter based on masking thresholds estimation in my development branch. This is better to avoid musical noise than using a postfilter because not only filters musical noise inbetween phrases or notes but inside them too. Most commercial tools like izotope spectral denoiser and others have this too. Besides I've been experimenting with transient protection conditions to avoid softening transients in musical material. It works quite well too. I had talked to Steve from audacitys team and I offer myself to improve current audacity noise reduction tool. That should tell you how well it works.