Spectrographic analysis of speech is one of the most widely used techniques for studying the acoustic-phonetic characteristics of different sound units in a language. It is an extension of the short-term spectral analysis, and primarily involves representation of the 3-D spectral information obtained by computing the magnitude spectrum over short overlapped window segments, i.e., 2-D spectral content varying with respect to time. The 3-D spectral information is represented on a 2-D plane with the X-axis representing time, Y-axis representing frequency, and the third dimension denoting the log-magnitude of the sinusoidal frequency components is converted to a proportional intensity of gray value. The resulting representation is referred to as a spectrogram.

Two popular spectrographic representations used for analysis are wideband spectrogram and narrowband spectrogram, depending on the spectral and temporal resolution preserved in the final representation in the frequency domain. In wideband (WB) spectrograms, the spectral information is averaged over frequency windows of bandwidths 100 to 200 Hz. The corresponding time window chosen is 5 to 10 ms, so as to maintain unit time-bandwidth product. An example of a wideband spectrogram computed with a time domain window size of 5 ms and a shift of 2.5 ms is shown in Figure 1(b).

Figure 1:
Spectrographic analysis of speech for an utterance "toast as usual"
(a) Speech waveform
(b) Wideband spectrogram
(c) Narrowband spectrogram

Similarly, a narrowband (NB) spectrogram is computed using narrow filter function in the frequency domain, which in turn corresponds to a longer segment length in the time domain. Figure 1(c) shows a typical narrowband spectrogram computed for time domain window size of 50 ms, which correspond to a frequency domain bandwidth of about 20 Hz. Wideband spectrogram provides better temporal resolution, while narrowband spectrogram provides better resolution in the frequency domain. The periodicity in the speech signal due to vibration of the vocal folds is reflected as vertical striations in the WB spectrogram. Horizontal striations in the NB spectrogram are due to the window size being shorter than pitch period, as can be seen from Figs.1(b) and (c), respectively.

The dark bands in the spectrogram represents the resonances of a vocal tract system for the given sound unit. These resonances are also called as formant frequencies which represents the high energy portions in the frequency spectrum of a speech signal. The shape of the dark bands indicates, how the vocal tract shape changes from one sound unit to the other.

1. Identifying the Voiced/Unvoiced/Plosive/Silence regions using spectrogram:

Voiced: In the case of vowels a regular formant structure (3 to 4 formant frequencies) and pitch harmonics (vertical striations in the case of wideband spectrogram) are used for identifying the voiced regions, where as nasals and voiced stops low frequency regions and pitch harmonics are used as clues.

Unvoiced: Energy at high frequency regions and no regular formant structure

Plosive: A silence bar followed by energy at high frequency regions.

Silence: No frequency components (white region)

2. Observing the time varying system and excitation characteristics using spectrogram

So far we described the sound units interms of acoustic phonetics, time varying excitation and time varying system characteristics. Here we demonstrate the time varying excitation and system characteristics (given in Table 1 and Table 2 respectively) using the spectrogram of a speech signal. The speech waveform, its transcription and NB spectrogram are shown in Figure 2. Table 3 presents the spectral details of different sound units using spectrogram.

Figure 2:
Speech waveform and its wideband spectrogram
for the utterance "kitAb mEj par hai"

Table 1: Time varying excitation characteristics from spectrogram

Table 2: Time varying system characteristics from spectrogram

Table 3: Spectral details for different sound units