254
QUESTIONS OF CREATION OF SYSTEMS OF SPEECH RECOGNITION
R.Mukhammadiev (NUU, student),
X.Shoqosimov (TUIT, student)
Known, it is perceived the auditory system of the person consisting of the ear, the auditory
nerve and the relevant Department of the Central system. The ear consists of three parts: outer,
middle and inner. The outer ear through the eardrum, connects
the middle ear with the
environment, where the distributed audio signal. The middle ear connects the eardrum with the
cochlea's internal membrane, and works as a lever system. Snail is a key element of perception of
sound waves, it contains the end of the auditory nerve. When sound excitation is the basic
membrane, consisting of several thousand fibers stretched across the cochlea, lymph of the cochlea
of the inner ear make the wave oscillation. On the one side of the membrane is the organ of Corti,
containing 20 thousand nerve endings of the auditory nerve, which
when excited are sent to a
Central system current pulses, the membrane along the cochlea .signal turns out. ACP performs
the job, which could be presented as two stages of process: 1) discrete; 2) quantization.
So the system works the sound of human perception, gifted by nature. When you create
systems for speech recognition we are not able to reproduce exactly the same system as the various
processes occurring inside the ear is still not mathematically justified. The human system of sound
perception so unique that it may sound to determine the location of the object, by the tone of the
speech to determine the internal state of the object. The speech signal has plenty of advantages,
which make it a promising medium for the transmission of information in various fields. However,
despite the many strengths of the speech, the issues of creation of artificial recognition systems
continuous speech and their practical application has been studied less extensively than the
technical basis, which could be used for this purpose. When you create a system of continuous
speech recognition there are many obstacles. Communication
from the point of view of
information theory is the process of transmitting information from source to receiver. In this
process, the speech signal has high redundancy (meaning only clean speech signal without noise).
Usually the speech signal contains noise and distortion. In this regard, there is a need for the
provision of speech signal from the noise, which is a separate and complex task.
The main principle of creation of systems of speech recognition is
based on the principle of
operation of the human ear and consists of three mandatory processes, which is shown in figure 1.
Fig.1. Major Subsystems of the speech recognition
It is known that the voice signal is analog. For voice signal processing in a computer, using
an analog-to-digital Converter (ADC) turns the digitized signal. ADC performs work, which can
be represented in the form of a two-step process: 1) sampling, 2) quantization.
In the first stage the signal S(t) is converted into a sequence S(n) =S(nT) where samples S(n)
are presented with unlimited precision. In the second stage, each sample S(n) is replaced with the
corresponding, binary number of trailing digits, you get CV(n) that operate when creating vector
features of speech. Because
it is a dynamic process, as OS the main features of speech, it is
impossible to use hangs from time to time. To calculate the characteristics of speech signals, you
must perform the following procedures: temporal processing; spectral analysis; perceptually
processing; the transformation parameters.
After all these calculations and analyses all the main data systemati-seroude and creates a
feature vector in the form X =(x1, x2,..., xn). In the process of recognition is the main work. There
Digitizing of a
speech signal
Calculation
Attributes
of a speech signal
Recognition
Speech signal
255
are many recognition algorithms of speech signal. The most simple and effective method of
recognition of the speech signal is compared with a finite set of pre-established governmental
standards and to determine the best fit. It is necessary to consider several factors: 1) different
implementations of the same words have different length; 2) in one implementation, the speed of
pronunciation of words may vary. From this it follows that the optimal alignment procedure model
and implementation must be nonlinear. Dynamic time alignment is
an effective method of this
optimal nonlinear alignment. There are many varieties of these algorithms. For example, they can
vary the local limits of ways by introducing different weights of transitions or other restrictions.
This algorithm is suitable to calculate the measure of similarity of the reference and
implementation by using the method of States. In this case, the vertical transitions are seen as
impossible, and the goal of the path search becomes the maximization of the likelihood.
Do'stlaringiz bilan baham: 
