I tried this new feature with 2 distinct voices (male and female) which overlap quite a bit. There was no noise in the background. The voices were separated only on very few parts. The “Voices Discrimination” parameter doesn´t seem to do anything, Voice 1 was always the same (I did a null test with 10%, 50% and 90%).
So I extracted the example from the SpectraLayers 10 World Premiere video (beginning at around 6:45). In the video the result is great but I can´t reproduce it. I still have a bit of overlap from the male voice and the separated male voice has artefacts. Both things are not existing in the video.
@Robin: Do you have an explanation?
1 Like
Can you send me your audio samples (contact [at] divideframe.com) so I can give it a direct try ?
I installed patch 1 and tried to separate the vocals from the SpectraLayers 10 World Premiere video again. But I get the same results and not the good ones from the video. Please let me know how the presenter (Mike?) in the video achieved that. Things in official product videos should be reproducible. Sorry for being a bit tenacious about that.
1 Like
I’m working on this in SL11…instead of starting a new thread, I’ll resurrect 
In my situation the recording on location in a windy vineyard in the south of France. Main contributor is wearing a Senn MKE2 Gold TX’d by a Senn EW100 to Sound Devices 722 and a shotgun Senn ME600 with just a Rycote dead cat. In the scene, the main contributor speaks to some un-mic’d passers-by and both the Lav and the Shotgun picked up the chatter. The unmic’d speakers are speaking at the same time.
I first ran Unmix Noisy Speech on each mic. On both, the unmic’d chatter went mostly to the Unmixed Noise layers.
Then I tried the Unmix Multiple Voices with no useable results…the module mis-identified some speakers. I’m not complaining. mind you…its just what I experienced.
OK, that said, this scene is important to the film so I set out to manually unmix and am getting amazing results.
To manually unmix, the harmonic selection tool is the first port of call; then transfer the remaining. If you can’t get the harmonic to work, might need to start with just a single freq. Using different colors for layers, you can fairly quickly create a “guide” layer and then extract each speaker’s contribution bit by bit and mix the voices to a pretty amazing result.
Yes, this isn’t automatic…it is taking days of post…but impressive results nonetheless
Yeah, uh, wow…creating a new balance of humans speaking from mono recordings and having control over the noise floor is an absolute pleasure! Takes a long time, but I’m seriously impressed
sadly, it is NOT automatic from a module at this stage
The unmixing levels feature is a workaround, you can use it to control between partials and noise levels, the key is to tinker around the fft window size (as that gives different results). However that process is not done in real-time which makes it difficult to determine noise floor from partials (because it wont allow to preview at such a high resolution). The horizontal (along with the vertical axis) is really where most of the processessing power seems to take place.
1 Like
I’m mostly using unmix levels on very severe speaker breakers- for which I’m getting fantastic results
the crunchy junk in a wireless lav (which can really annoy me, but typically not others) can get cleaned up really well with unmix levels IME.
But I’m not using unmix levels everywhere…it’s just adding too much complexity…that’s typically what my deeper nested layers are…and I’m struggling enough with the mute/solo functionality.
I certainly do not tinker with the FFT size very much…I’m mostly working at how I can best see the spectrograph which is 3072 smp/ Blackman-Harris…and directionally straight up/down