Unmix Noisy Speech leaves out chunks of obvious human speech

SpectraLayers
,
1

I really wonder why this happens… Every single time I use the Noisy Speech Unmix I always get bits and chunks of really obvious human speech that is sometimes detected if the file length changes, or will be detected by the likes of Waves Clarity VX or Acon Extract Dialogue. Sometimes the omitted bit (placed in the Noise layer) will be cut off abruptly from a correctly detected bit (placed in the Speech layer). I noticed this behavior with both CPU and GPU processing. My version is 10.0.30

2 Likes
2

Can you send me one or two samples that shows such issues ? Either upload it somewhere and PM me, or to contact [at] divideframe [dot] com.

1 Like
3

I have the same situation. I am currently able to work with the dropouts, selecting and copying from the Noise layer to a third layer. The unmixing was mostly very good. My main character sounds like himself for a change!

1 Like
4

I know this is a year old thread, but I decided to add to this post rather than start a new one. Currently on SL11, this is one of my biggest time soaks. Thanks to the Transfer tool, moving these speech remnants from the separated noise layer back to speech layer or to new layers is clearly addressed…nevertheless this process is very time consuming.

In my situation, I can often have multiple speakers at the same time and Unmix Noisy Speech does incredible separation, no doubt…a massive time saver nearly impossible for a human to carry out and no way could human carry out that separation work as fast as running the module. I just thought I’d share some screen shots.

Source: human speech on location (outside) Sennheiser MKE2 Gold>ew100 g Wireless> Sound Devices 722
(comped in SAW Studio with Acon DeClicker and Fab Filter Pro DS and no other processing other than some unwanted wind and some sync fx moved to other layers)

Further, this is two people talking into the same mics, yet it would be pretty much the same separation with just a single speaker.

Original file:

Unmix Noisy Speech module Speech layer

Unmix Noisy Speech module Noise layer

Unmix Noisy Speech module Speech (magenta) + Noise (cyan) layer

Unmix Noisy Speech module Speech (magenta) + Noise (cyan) layer zoomed out a little to show the next phrase

Just sharing my experiences here. I have found subsequent passes of the Unmix Noisy Speech module on the Noise layer does not separate much more these speech remnants…I just get busy with the Transfer tool.

1 Like