Parameters choice
Heisenberg’s uncertainty principle
A signal cannot be both well localized in time and in frequency.\
Consequence: short window are more adapted to “transient”, and long window to “tonal”, “stationnary” part of the signal
Common choices
For a high-fidelity audio signal with sampling frequency at $44.1$ kHz, with an analysis window of size $L$.
- Shape of the window: Hann, Hamming, Gaussian
- Length of the window: between 256 samples to 4096 samples. Common choice: 1024 samples
- Redundancy in time: overlap of 50% or 75% between two consecutive window
- Redundancy in frequency: FFT of size $L$ or $2L$
Useful toolbox
- Matlab: The Large Time-Frequency Analysis Toolbox (LTFAT)
- Python: LTFATPy (small part of the matlab toolbox)