pico VS slttblep

The landmarks extraction binary models are already available. The only missing part is the C or Javascript code to load the model, and perform landamarks extraction given a face coordinates. Landmarks extraction code is based on the Pico JS face & pupil detection library at: https://github.com/nenadmarkus/picojs which is in turn based on the Pico C library by the same author: https://github.com/nenadmarkus/pico

The oscillator bit is, but I don't appear to have uploaded the one with the filter which I will need to find.

https://github.com/ErroneousBosh/slttblep

The filter is just an SVF with, with a precomputed expo scale "bent over" at the top to correct for quantisation. From that the two SVF coefficients ω/Q and ω*Q are calculated every time there's a control update and of course because you can't divide on an Arduino it uses a lookup table of reciprocals. I could probably use a "wider" table of 16-bit values for better precision.

This code here will run on an Arduino and read an analogue input which sets the pitch of a bandlimited sawtooth waveform coming out one of the PWMs.

Take a look at https://github.com/ErroneousBosh/slttblep for an example of an anti-aliased sawtooth on an Arduino. Since you're using an 8-bit Arduino you can't really multiply and can't divide at all, so you have to use the titular Stupid Lookup Table Tricks to "divide" by multiplying by a reciprocal, and use a precomputed table for the function that "smooths out" the step.

I've waffled long enough. Get an Arduino of some sort, and run this on it and read through the code until it makes more sense :-)