Zen
                                
                                 Zen copied to clipboard
                                
                                    Zen copied to clipboard
                            
                            
                            
                        Add a CQT
Applying median filtering on a CQT with 96 bins per octave gives a better separation quality - and it should still be achievable in real-time.
https://www.univie.ac.at/nonstatgab/slicq/index.php
NSGT-based realtime, invertible CQT is my preferred choice here.