Short-Time Fourier Transform

Window length determines the time-frequency trade-off. A rule of thumb is that you need at least 2-3 cycles of the lowest frequency of interest. For music, 512-2048 samples is typical; for speech, 256-512 samples. Shorter windows preserve temporal events; longer windows provide better frequency resolution. Experiment with different lengths and visualize the resulting spectrograms to find the optimal balance for your application.

Hann and Hamming windows are good general-purpose choices with reasonable trade-offs between main lobe width and side lobe levels. Blackman windows have lower spectral leakage at the cost of wider main lobe. Rectangular windows should be avoided unless you specifically need frequency response peaks. Choose based on whether you prioritize frequency resolution (wider main lobe acceptable) or spectral leakage reduction (lower side lobes needed).

50% overlap is a common baseline that provides smooth time evolution without gaps or redundancy. For synthesis applications like phase vocoder, 75% overlap (hop size = window/4) often works better. Higher overlap increases computational cost but improves temporal smoothness. Lower overlap reduces computation but may create discontinuities. The optimal choice depends on your specific application and processing needs.

A spectrogram is the magnitude-squared (power) visualization of the STFT. Other time-frequency methods like continuous wavelet transform and Wigner-Ville distribution exist but have different trade-offs. Wavelets adapt resolution to frequency (lower frequencies get better frequency resolution), while STFT maintains fixed resolution. STFT is simpler and computationally more efficient; choose wavelets if you need frequency-dependent resolution.