|
Main directions of scientific research |
|
|
|
Scientific publications about speech technologies |
|
|
|
Computer demos and presentations |
|
|
|
Contact information (phones, e-mails) |
|
|
|
Information on Basic Directions of Scientific Research
The basic work of our group of mathematicians is
devoted to theoretical problems in automata theory. An automata is a device that has input and
output channels and internal memory and is capable to translate input
influence into output reactions. It
is possible to get new automata using the existing ones by connecting the
input and output channels. A complex analysis is required to find the
principal possibility of solution of this problem (that is called the problem
of completeness). The
problem of completeness is falls under the category of problems unsolvable
in general case. It is possible to solve it for special classes of automata.
The mathematical apparatus that arises from here is a set of rules
of input information tests. It is possible this apparatus in an applied investigations on
recognition of consecutive information.
As a result of this, it become possible to construct
recognition algorithms and models of artificial organs of sense in
parallel with the theoretical work.
Our main scientific works are:
- 1985 - model of color eyesight of humans and animals (in co-operation
with the chair of psychophysiology of the MSU);
- 1987 - an automata algorithm of photo pictures recognition;
- 1993 - semantic frequency analysis of texts;
- 1994 - the management of all-Russia speech recognition project;
- 1995 - model of subjective phoneme perception;
- 1996 - the problem of separation of symbols in printed texts and
hand-writing;
- 1998 - usage
of non-audio sensors for speech recognition;
- 1999 - lip reading, hand-writing recognition, the model of constancy of
color eyesight of humans.
The work on all these problems were done in the form of
students diploma works, each including a theoretical part and a demo
computer program. They used
an automata model of input information representation.
The
theoretical investigation on perspectives of usage of new mathematical
methods for speech recognition can be fulfilled in the following
directions:
Signal level.
- Search for new types of speech signal
parameters. Reduction of dimension of speech signals description
space using special methods like factor analysis and others.
- Making experiments on algorithmization of ways
of speech recognition by humans. Investigation of hearing apparatus
neurophysiology. Construction of subjective sound metrics.
Finding the dimension of space of subjective speech features.
Testing the hypothesis of local speech recognition by humans (in
co-operation with the chair of psychophysiology of the MSU)
- Search for speaker-independent features on
reliable phoneme segmentation of speech. Investigation of
possibility of local phoneme identification. Investigation of
dependence of sounds parameters on phonetic environment. Automatic
extraction of features, allowing us to define the class of phoneme.
- Investigation of possibility of using the
additional (non-acoustic) sources of information for speech recognition.
Semantic level.
- Automatization of process of construction of
semantic grammars in the given theme.
- Investigation of possibilities of semantic
grammars usage. Construction of themes library.
- Algorithms of fast recognition of theme of
speech using audio information from the known list of possible themes.
Mathematical methods of recognition.
- Search for parameters of stochastic vectors,
characterizing the phonemes. The investigation for type of its
stochastic distribution. Modifications and enhancing of HMM.
- Construction of discrete methods of fast
pre-recognition on the base of phoneme segmentation and using the features
of separate phonemes.
- Syntheses of hierarchical recognizer -
structural automata, working with different levels of details of speech
signal simultaneously.
- Investigations of possibilities of usage of
methods of fussy logic, optimal control and others in solving of problems
of speech recognition.
- Usage of autocorrelation and autocomparing
methods for solving a problem of speaker-independent recognition.
Deputy
head of the Laboratory of Problems of Theoretical Cybernetics of the Lomonosov Moscow State University
Prof.
Dmitry
Babin
|