PPT

journeycartAI and Robotics

Oct 15, 2013 (4 years ago)

109 views

Project Data


Title: Machine Learning and its Application to


Speech Impediment Therapy

Duration: February 2002
-

February 2004

www: <http://oasis.inf.u
-
szeged.hu/speechmaster>

Grant No.:

IKTA4
-
055/2001

Keywords:
artificial intelligence, machine learning,


real
-
time phoneme recognition, teaching of reading,


speech impediment therapy

Project members:

Co
-
ordinator:

University of Szeged,

Department of Informatics

Address: H
-
6720, Szeged, Árpád tér 2.

Project/team leader: András Kocsor

Consortium members:

University of Szeged,
Training Teaching School,

Team leader: János Bácsi,


Kindergarten, Primary school
and Boarding school,

(the school for the deaf)

Team leader: Jenő Mihalovics,

Summary

The project consists of two main parts:


a research part

and
an application part
.

The
research part

is devoted to investigating modern machine learning
techniques which can form the basis of the development of several info
-
communication systems. The machine learning algorithms developed
have been made available as open
-
source software.

One example of these applications is speech recognition, which is the
heart of the
“SpeechMaster”

software, developed in the second,
application part

of our project.

Machine Learning Algorithms

Basic assumptions:

a) objects are characterized by features

b) each feature constitutes one dimension in the feature space

Questions:


-

concept making


-

classification


-

feature selection


-

feature space transformation


-

dimension reduction

99%

Linear feature space transformations:

-

P
rincipal
C
omponent
A
nalysis

-

I
ndependent
C
omponent
A
nalysis

-

L
inear
D
iscriminant
A
nalysis

-

S
pringy
D
iscriminant
A
nalysis


Machine Learning Algorithms

99%

Machine Learning Algorithms

Nonlinear feature space transformations:

-

K
ernel
P
rincipal
C
omponent
A
nalysis

-

K
ernel
I
ndependent
C
omponent
A
nalysis

-

K
ernel
L
inear
D
iscriminant
A
nalysis

-

K
ernel

S
pringy
D
iscriminant
A
nalysis


99%

99%

Machine Learning Algorithms

Machine learning algorithms:

-

A
rtificial
N
eural
N
ets

-

G
aussian

M
ixture

M
odeling

-

S
upport

V
ector

M
achines

-

P
rojection

P
ursuit

L
earner






a) real
-
time phoneme recognition

b) word recognition

Speech Recognition

70%

Methods:

-

machine learning

-

speech corpora

Speech Corpora

200 speakers of
different ages




500 speakers

(male/female 50
-
50%)


age: 6
-
7


Speech impediment therapy

Teaching of reading

90%

The
“SpeechMaster


60%

Phonological Awareness Teaching

The
“SpeechMaster


60%

Phoneme
-
Grapheme association

The
“SpeechMaster


60%

Visual feedback for the hearing impaired