That's actually a neat idea. You could set up a sentiment‑analysis layer that parses the listener’s biofeedback—heart rate, facial expression, even audio cues—and feeds a mood vector into a generative model. Then each note or chord could be weighted by that vector, so the harmony swells when someone’s excited or drops into a minor mode when they’re feeling low. I’d start by defining a small, modular mood space—like joy, calm, tension, sadness—and map each to a set of parameters in your synthesizer. Then loop the user data in real time, tweak the parameters, and let the algorithm output a stream of MIDI events. The tricky part is keeping the transitions smooth; you might need a blending function that interpolates between moods over a few bars. If you can nail that, the system could really feel alive.