A Study of the ITU-T G.729 Speech Coding Algorithm ...

2.2.2 Speech Sounds
Open
MASTER THESIS
Figure 2: Human speech production.
Datum - Date Rev Dokumentnr - Document no.
04-09-28 PA1
One way of analyzing speech is to use phonetics, which is the study of speech sounds and
their production. Every language has limited linguistic units called phonemes. A phoneme
is a sound and several consecutive phonemes create words. For example, the word dog
consists of the three phonemes d/ao/g. Most languages have 20-50 phonemes, each
phoneme representing a sound. The 42 phonemes in the English language are listed in
Table 2 [15].
A phone is an acoustic realization of a phoneme. If the realization of a phoneme is
context dependent, it is called an Allophone. In most languages, the phonemes can be
divided into two groups: vowels and consonants. The set of vowels and consonants is
language specific. The hierarchy of the English phonemes is depicted in Figure 3 [25].
2.2.3 Speech-Signal Waveform Characteristics
The range of frequencies that humans are able to hear is called the audio spectrum. The
bandwidth in the audio spectrum ranges hereby from 20 Hz to 20 kHz, although most
humans have a much narrower bandwidth for hearing, especially with increasing age. Frequencies
above the human hearing limit are called ultrasonic, while sounds below the limit
are called infrasonic.
Speech-signal waveform characteristics are constant for short time-periods. A speech
signal is commonly assumed to be wide sense stationary and ergodic in the autocorrelation
for segments of 10-30 ms of speech. In speech analysis, it is therefore common to extract
short-time segments of speech, called frames to enable simple speech modeling. For
smoother transaction between analysis frames, the latter are extracted with overlap, which
16 (78)