Noise Reduction - tips and tricks
Posted: Jul 02, 2024 4:48 am
The whole field of noise reduction has changed dramatically over the past few years with the introduction of machine-learning algorithms that can routinely perform miracles.
I've posted elsewhere about my experiences with Hush Pro, which I demoed to professional sound recordists earlier this year. Without exception, their jaws dropped to the floor on a demo of a personal lav mic where the dialogue was drowned out by seawash. Hush Pro was essentially perfect, separating the voice and seawash into separate files with no discernable quality degradation at all. It really did feel miraculous.
I recently tried it on a job even though the main rig is Windows and this is Mac-only. I figured I could render files on the Macbook, and transfer back. This particular job was recorded in a kitchen - plenty of room noise, sizzling pans and various cookery noises that I needed to separate into voice and effects files. To my unhappy astonishment, Hush Pro didn't detect any of the cookery noises at all, even the sizzling, and kept it all as dialogue (it did a pretty good job on the room reverb though). It really hit home to me just how variable these tools can be - the same plugin that produces perfect results on one clip can not even touch another. At the moment it feels like the more plugins you have the better your odds.
In terms of more forensic work, one trick I've used possibly more than any other over the years is the most basic - copy and paste. So often the best and cleanest solution to a specific nasty problem is to copy a part of the frequency band next to the problem area, and paste it over the thing you don't want, a la photoshop. Any editor with a spectrograh view can do this - I use RX.
I haven't yet found any plugin for a job like mouth delick or deplosive that is 100% safe to use in auto-mode - even set conservatively they can falsely detect real speech frequencies and eliminate, or it can miss stuff. RX's tools here are very good but not infallible, so I just do it when I need it (and I often just choose the rough frequency area with the problem, rather then do it for an entire broadband file). For deplosive, I sometimes do a different very lo-tech trick - especially for female voices I just use a 48db high pass filter. I look on the spectragram for the lowest genuine speech frequency I can see and set the HPF to right below this. It's amazing how well it can work - in these cases I have been able to blat an entire batch and forget it.
I've posted elsewhere about my experiences with Hush Pro, which I demoed to professional sound recordists earlier this year. Without exception, their jaws dropped to the floor on a demo of a personal lav mic where the dialogue was drowned out by seawash. Hush Pro was essentially perfect, separating the voice and seawash into separate files with no discernable quality degradation at all. It really did feel miraculous.
I recently tried it on a job even though the main rig is Windows and this is Mac-only. I figured I could render files on the Macbook, and transfer back. This particular job was recorded in a kitchen - plenty of room noise, sizzling pans and various cookery noises that I needed to separate into voice and effects files. To my unhappy astonishment, Hush Pro didn't detect any of the cookery noises at all, even the sizzling, and kept it all as dialogue (it did a pretty good job on the room reverb though). It really hit home to me just how variable these tools can be - the same plugin that produces perfect results on one clip can not even touch another. At the moment it feels like the more plugins you have the better your odds.
In terms of more forensic work, one trick I've used possibly more than any other over the years is the most basic - copy and paste. So often the best and cleanest solution to a specific nasty problem is to copy a part of the frequency band next to the problem area, and paste it over the thing you don't want, a la photoshop. Any editor with a spectrograh view can do this - I use RX.
I haven't yet found any plugin for a job like mouth delick or deplosive that is 100% safe to use in auto-mode - even set conservatively they can falsely detect real speech frequencies and eliminate, or it can miss stuff. RX's tools here are very good but not infallible, so I just do it when I need it (and I often just choose the rough frequency area with the problem, rather then do it for an entire broadband file). For deplosive, I sometimes do a different very lo-tech trick - especially for female voices I just use a 48db high pass filter. I look on the spectragram for the lowest genuine speech frequency I can see and set the HPF to right below this. It's amazing how well it can work - in these cases I have been able to blat an entire batch and forget it.