Voice masking for removing voice and not birdsong

Question

First the problem I'm tyring to solve: I'd like to find a way of removing any potential noise from audio files recorded in noisy environments where we are trying to record birdsong. This is for the sake of privacy, we cannot record anything unless we're sure we aren't recording private conversations.

So, what is the potential for voice masking in this problem? I've done research but it is hard to understand the real potential without learning a lot of background science first. So I'm hoping someone in the community can help.

Basically I'd like to know if voice masking techniques are able to be used on a random sound file to mask all potential voice so it is not understandable... at the same time as ensuring that we aren't masking other sounds, especially birdsong. Is this doable?

Answer 1

Sound masking in time-frequency is a very complex problem, especially to do it in real-time while collecting data. For ensuring privacy I would suggest using a simpler approach.

Sparse random sampling

Instead of recording continuously for hours (potentially picking up whole conversations), record only short segments spaced from each-other in time. They need to be long enough to capture events in question (birdcall/birdsong), but short enough to not be able to capture much coherent conversation.

For instance record 0.1-1 second segments every 10-100 seconds. This can be done on a fixed schedule, or using reservoir sampling.

This approach has the advantage of not introducing any bias in the selection. It reduces the amount of data, and limits analysis to short time windows. It may also meet the privacy protection requirements.

Drop samples with Voice Activity Detection

While recording, consider segments of 1-60 seconds. Run a standard Voice Activity Detection (VAD) algorithm on each of them. If voice is detected, discard the entire segment. Need to use a low threshold for the VAD to get a low amount of false negatives (sample actually has speech).

Sampling only when the VAD does not trigger may introduce bias in the data. For instance if there is humans talk at certain times of day. Or if some birdsong is confused as human voice by the VAD.