Speech Signal Analysis and Modelling - Digital Speech: Coding for Low Bit Rate Communication Systems - page 61

(b)

(a)

0.0 55.0 110.0 165.0 220.0 0.0 1.0 2.0 3.0

Time (samples) Frequency (kHz)

(c) (d)

0.0 55.0 110.0 165.0 220.0 0.0 1.0 2.0 3.0

Figure 4.5 Effects of window types on unvoiced speech with a 220 sample window
length: (a) and (b) show the time and frequency plots of speech using a rectangular
window, and (c) and (d) show the time and frequency plots of speech using a
Hamming window

In order to see the effect of varying thewindow length, consider the example
in Figure 4.6 where a block of 40 sample (5ms) long voiced speech is shown.
In this case, the time domain speech s(n) w(k n) shown in Figures 4.6a and

4.6c do not show the signal periodicity accurately. This is also true for the
signal spectra shown in Figures 4.6b and 4.6d. When compared with Figure
4.4, the spectra of Figure 4.6 show only a few rather broad peaks at about
500, 1350, 2300, and 3400Hz corresponding to the formants of the speech
contained within the window.
The effects of Hamming and rectangular windowing are still visible in
the spectra of Figures 4.6b and 4.6d. If windows of 5ms duration were to
be positioned at the beginning and end of the 27.5ms interval they would
show different spectral characteristics. Therefore, good temporal resolution
requires a short window while good frequency resolution of speech requires

−