The typical window length is 20ms. For 10 kHz sampling frequency, 200 speech samples are used, padded with 56 zeros, hamming windowed, FFT and converted to a power spectral density.
