I'm working on a project where the purpose is to enhance FFT peaks in a signal (musical instruments), by other words, reduce non-harmonic components and enhance the identified FFT peaks. So far I'm doing the following:
- I'm analysing the file on a frame basis and
- for each frame, I calculate an AR(8) model,
- then reduce the bandwidth for each peak and
- apply this as a filter to the frame,
- combine everything using PSOLA
The results are good but they could be better. Any ideas for a different approach?