Realtime score: how to measure similarities of 5 different live sources?

Getting Started and Wayfinding
1

Cheers, newbie Q here:
I’m preparing a realtime score in SuperCollider for 5 instrumentalists that I want to base on measures of similarity between the mic’ed signals of the performers.
(E.g. clarinet playing similar to flute, gets instruction “cl play less similar to fl!”
Or, if A sounds very different from B, tell A to play more like B)

After fuzzing too long with SC’s rather cumbersome Machine Listening tools, I reckon this is a good place to get into FluCoMa to do the job.

Can you sages advise me how to proceed at calculating differences/similarities between 5 real time signals?
As a minimum of analyIs features, I thought of the spectral shape descriptors and the rate of Novelty triggers.
Averaging over a short term of 0.5 - 3 secs shd suffice; sending to client a few times a sec; then doing the decisions + display from there.
It’s not clear to me, though, how to go about reducing the different delta values of the live signals to a single difference measure.
Could you give me some hints?
Very appreciated!
Thanks,
Hannes

2

Your plan sounds like a good starting place. Maybe this article on scaling will be useful:

why scale?

FluCoMa doesn’t have any built in functions that only ask “What is the distance between these two vectors?” but these are generally really straight forward to compute. More on that here:

https://www.analyticsvidhya.com/blog/2020/02/4-types-of-distance-metrics-in-machine-learning/

1 Like
3

If we really want to geek out here, and maybe this is more than the OT is asking for… kmeans would give you the distance to each ‘cluster’ so if the machine listening part (the numerical representation of each ‘class’) is segregated enough to be clustered in 4, then getting the distance from each centroid would be available there…

1 Like