Eliminate voices/talking from recordings? (1 Viewer)

[schwartzie]

Hello,

I'm seldom awake early enough to catch the best of the dawn birdsong, so I decided to set up a fixed recording installation so I can replay it later in the day. I rigged a stereo mic under the roof eaves, and connected it to a Raspberry Pi via a USB audio interface. The Raspberry Pi records continuously, capturing 5-minute files and uploading them to a private cloud account where I can listen to them and process them with BirdNET-Analyzer.

It's been working nicely as intended, but an unwanted side-effect is that I've created a powerful surveillance tool that captures many of my conversations outside and even inside through open windows

I tried applying high-pass filters with Audacity and SoX at ~400Hz, but even with multiple passes and aggressive roll-off settings, speech seems to be only attenuated and not eliminated. Based on this reply to a related thread, I suspect filtering won't be enough here: Human cancelling headphones.

Is there anything else that might work to eliminate or better attenuate speech in these recordings as I capture them or in post-processing?

Thanks!

 

[Jon.Bryant]

The human voice has a frequency range from about 125hz to 8khz, so the sound covers the same frequency range of many/most birds. Also the human voice is wide spectrum - by this I mean that a human voice creates sound simultaneously over a wide frequency band. This means that if you record speech at the same time as a bird vocal, these is a very high chance that the human voice will coincide with the frequency range of the bird. This means you cannot really apply a filter or equaliser - filters and equalisers work by reducing sound at selected frequencies, so only work if you have a relatively narrow band sound on a completely different frequency to the bird vocal - say low frequency wind noise.

Removing human voice is tricky in post, as you are generally looking to disentangle two overlapping signals. This cannot be readily achieved by a standard function. Some products have started to include AI functions, that can separate overlapping sounds, but as you can imagine the AI engines are not designed to separate bird sounds from human voice - a pretty niche requirement. The engines at the moment can deconstruct music into vocals and various instruments, or separate sound on general shape of the sound or loudness. If you separate sounds on shape (tonals, transients or noise) or loudness, you may then be able to manually tease apart the bird and human vocals, then delete of reduce the latter. The problem is that software with this capability is not free and that the process would be time consuming - the AI processing takes some time, but the subsequent manual editing a lot longer. I therefore only venture down this route for the occasional ‘special recording’ which is worth trying to salvage. From the description of your setup, I think applying this fix to all recordings would be far too burdensome and labour intensive.

I therefore think there is not currently a solution to your question - unless you are willing to impose a silent hour at home!

 

[Mike Crawley]

low tech solution ?
Put the recorder on a timer.
Start just before the Dawn chorus and off at the approximate time you awaken each morning.

 

[schwartzie]

@Jon.Bryant thank you for your detailed and informative reply! You've saved me a lot of fruitless toil here and I appreciate your expert guidance in helping me reframe my objectives for the project.

@Mike C I could but then I'd miss other exciting action like the nighttime howling of the coyotes I agree though with your premise of making the "surveillance" less invasive/problematic, and will likely end up going in a direction like it.

Thank you for your interest and suggestions!