Level-wise they are separate though. It should be easy enough to find the proper gain for the reference recording (how loud does it sound in the video?), and then subtracted, leaving any residual energy in that band from desired sources. Yeah, there will be some gain-matching and spatial artifacts, but probably not enough to make things (voices for instance) unintelligible.

Thats actually where any ML component would naturally sit. It would try to basically determine a "close-enough" deconvolution envelope for the reference recording.

No, but it will preserve them online for later filtering/post processing.

Maybe, depending on the details of your scheme, I guess? If you subtract data, that data is gone, and there's no separate stream of data in the recording for the cops' voices vs. the phone.

We just "play" the song - so it's a spotify call. Ideally anything not aligned with the song meant to induce takedown and the inverted playing of it will still be maintained.

To clarify, we don't want to subtract the data from the feed, we can just nullify it near the microphone.

I think it would be harder to filter on an instagram feed directly, not sure how their api looks or if it's possible?

I see, you're saying play the inverse at the recording time. So I misunderstood, I thought you were meaning doing it to the recording.

This is also complicated, because you need to factor in the relative locations of the original source, your source, and the pickup to make sure the phase is correct.