Editing with a learned sound "signature"?

I am somewhat of a SpectraLayers novice, but have been amazed by some of the things I have been able to do with it.

One of my favorite experiences was with a long crappy shotgun recording of a baby sea otter. The otter was quite far away, and there was a ton of background noise: cars, boats, bells, chattering people, dogs, seals, machinery and other wharf noises. Nine minutes of loud boring din with just a few seconds of otter (baby otters, alas, do not perform on request).

The baby otter had a keening cry, with lots of upper harmonics. By fiddling with the amp/res/FFT parameters, SpectraLayers made it easy to visualize and highlight the otter cries, and spot every one of them distinctly in the nine-minute long recording.

Using the wand/harmonic selection tools, it was then a fairly quick process to separate and layer one of the cries, so that one layer had all the background noise with no detectable trace of the otter, and one layer contained a pristine (if distant) otter cry. With the manual tools, the separation was quite amazing.

One abstract result of the layered separation was a clear visual sense of a sonic signature: what a baby otter cry “looks like” and what gets left behind as background.

Is there any way to have SpectraLayers “learn” a sound signature from one or more manual separations, and then automate the process of finding and separating the rest of them in an audio file? (It would be a bit like the built-in Unmix tools, except based on very specific learned examples and presumably more precise: a kind of fuzzy logic find-and-separate.)

Baby otter cry:

1 Like

@Robin_Lobel, if this is not currently possible, is something like this on the roadmap?

1 Like

Have you tried “Select similar”?

1 Like

You know, I remember seeing it and trying it once in v9 (based on a video tutorial that matched kick drums), but at the time it didn’t behave as expected and I just kind of ignored it ever since. But now that you mention it, I wonder if that could work for matching complex, variable organic sounds from a noisy background like this.

A quick try in v10 did not do anything, but I need to play around more with how that function works with layer-split sounds. The “signature” cry was already meticulously separated from the main file into its own layer. Just like the previous image, that layer is mostly emptiness, with just the isolated sound. Maybe it involves somehow re-selecting this within the one layer without selecting the empty spaces, and then finding it in the other layer.

Will keep trying and report back. Thanks for the thought!

1 Like

The current pattern match will look for sounds that are almost identical to the one selected.
This will likely be extended in a future version of SpectraLayers, but there’s no precise roadmap for that yet.