In the basic implementation of CQT the width of temporal kernels depends on the central frequncy (higher frequency -> narrower kernel). The matrix of temporal kernel for a single time-domain signal frame for real-time usage contains kernels of various width aligned to the end of the frame. For an efficient computation we then use a matrix of spectral kernels.
Although adjacent frames may overlap high-frequency kernels cover just a small parts of each signal frame and low-frequency kernels cover highly overlapping parts of the signal. In high frequencies this lack of information causes significant noise.
- Would it be possible to compute with each kernel separately so that there would be denser coverage in high frequencies and sparser coverage in low frequencies?
- Does NSGT or Reassignment solve this problem better?