Ўзбекистон республикаси ахборот технологиялари ва коммуникацияларини ривожлантириш вазирлиги муҳаммад ал-хоразмий номидаги

SEGMENTATION OF SPEECH SIGNAL PROBLEMS 
F.Nuraliev (TUIT, Dean of the Faculty), 
D.Ibragimova (TUIT, student) 
Speech is the most simple and natural means of giving commands and information 
exchange for human. However, the problem of automatic recognition and speech perception is 
very difficult. This is due to the fact that the development of a model of the communication channel 
representing the interaction of the transmitter with a passive receiver, and it is insufficient for this 
task because: 1) the receiver needs to be active in the process of exchange of information using 
speech technologies; 2) some adaptation of the transmitter (speaker) to receiver (listener). 
Meanwhile, the speech signal has a number of advantages that makes it a very attractive medium 
for the information exchange with various technical systems. 
It is known that speech is represented as a continuous signal, where the limits of both words 
and their individual components are difficult to define with objective methods. It should be 
emphasized that between the auditory perception of speech (with its natural recognition) and 
speech perception (natural language understanding) in technical systems there is a huge difference, 
in particular, the boundaries between words, such seemingly clear under natural listening, not so 
obvious when analyzing the speech signals in technical systems. Between words, there are no 
special markers limitations. They are absent between the individual constituent parts of the word. 
Therefore, the issues of development, improvement and application of algorithms of segmentation 
of continuous speech signal are a very important task. 
The analysis of literary sources shows that speech is created in an almost continuous state 
change of the vocal apparatus during the transition from pronunciation of one sound to the 
pronunciation of the other. Conditional recording of the speech sounds is normally consists of a 
chain of atomic elements, called phonemes. For this continuous representation of the speech signal 
requires a set of acoustic speech sounds to be split into segments, which gives the opportunity to 
associate with specific phonemes. People are able to perform such segmentation, although 
depending on the cultural level of the same speech, they are segmented differently. Automatic 
segmentation of the speech signal can be carried out in the time and frequency domain or in time 
– frequency (spectrograph) [1]. Existing methods of automatic segmentation are mostly designed 
for each specific task. This report describes the segmentation of speech in the time domain. 
The essence of the segmentation of the speech signal method is as follows. First the sound 
wave is split into a number of elementary segments of the original and is represented by vector of 
amplitudes of the sound wave a=(A1, A2, ..., am). The length of each segment m is chosen small 
enough that there should not be significant changes in its limits. Typically, each basic segment 
contains about 50-100 points, separated by an interval of 80-100 µs. For each initial area of the 
speech signal is determined by:
• the maximum element (x1) of the components of the vector a in absolute value; 
• the number of zero-crossing (x2), which is calculated by the formula 

---

256 
,
1
1
2




m
i
i
b
x
where
b
i 
= 






.
o
,
0
;
0
i
,
1
1
therwise
a
a
f
i
i
Further, the original basic segments with similar characteristics are grouped into larger 
segments. It is known [2] that the intensity of all vowels, fricative, smooth, nasal sounds preceding 
the pause stops the greater part of the interval of sounding of the phoneme is nearly constant. 
Meanwhile, variations are observed only at the boundaries of phonemes. Analysis of computer 
experiments on segmentation of speech signals shows that x1is changed more than x2. Therefore, 
the original grouping of the elementary segments with similar acoustic characteristics can realize 
mainly by setting x1 and parameter x2 can be used to eliminate possible ambiguities.