(1/2)
I found a nice improvement to the Constant-Q Sliding DFT¹.
Using single precision calculation (32-bit float) generates some remanent noise because of the recursive nature of the algorithm: a running sum combined with complex rotations.
Replacing the complex multiply in eq. 3 for the twiddle factors (outer exponential) with the Martin Vicanek’s quadrature oscillator² helps a lot in the low frequencies. About 30 dB of improvement here, almost no extra CPU cost.
¹ A transform turning a PCM signal into a frequency spectrum. https://www.dafx.de/paper-archive/details.php?id=QMFQa1rIAM7mwu9tMIZAlg














