Вы находитесь на странице: 1из 1

Proceedings of The Arst Joint BMESlEMES Conference Serving Humanity, Advancing Technology O d 1318, 93, A M $ , GA.

USA

PROCESSING AND CLASSIFICATION OF DEFORMED SPEECH USING


NEURAL NETWORKS
RTadeusiewicz', A Invorski', W. Wszdek**,T. Wszdek**
of Mining and Metallurgy, Dept. of Automatics, Krakow, Poland
'University
it
University of Mining and Metallurgy, Dept. of Mechanics and Vibroacoustics, Krakow, Poland
Abstraei
In many problems of medical diagnosis as.well as thmpy
and rehabilitation, evaluation of the deformed speech signal
quality is required In the problems of dist0Ited speech
diagnosis the regular m e t h d of speech signal prep.ocessing
and classificatios, usedin speech or voice recognition, toraUy
fail. Also the standard speech signal parametriZaton
techniques (e.g. LPC or cqxtral ccefficients) cannot
satisfactorily describe the pathological speech because of its
dissimilar phonetic and acoustic structure comparing to the
c o m t speech, and also because the aim of the recognition
process is totally different. In the paper new methcd of the
processing and classication or pathologically deformed
speech, based on the neural networks techniques are
presented and discussed
Keywords: neural networks,pathological speech.
Introduction
While in the lypical speech rewgnition stndies the goal is to
reveal the semantic aspects of the pronounced text, in the
tasks of medical diagnosis based on the speech signal
analysis the semantic contents of the utterance is irrelevant,
and the required signal featores must be extremely sensitive
even to minor deformations in the layer related to the voice
structure and fonctioning and to the -shuchue of the vocal
tract. In the pracent study special attention has been focused
on the sttuchne .evaluation for the feature space demibing
the pathological speech si@ and its comparison to the
featnre space typically considered in the speech processing
problems.
Methods
The registration and -sing
of the speech signal has
been carried mt in an anechoic chamber using a specific
measurement setup, designed in the course of .@ow
research work The multispectra obtained from the analysis
has been transformed to wen dimensional vectors of
features X of the following form:

<M,,M,,M,, WS,, WSt, WS,, WS,>= X

where: MO, MI, M2, spectrll moments,


WSi - relative power coeffiaent, describing the ratio of the
signal power in the i-thband to the signal power in the whole
band (i = s, 1,2,3).
MDinresolts
In the class rewgnition for the speech signal pathologies
there are no f h d discriminates for the classes of deformed
speech ptIerns, and what's more it cannot be found in
advance how many classes the study will be able to

0-7803-5674-8/99/$10.000 1999 IEEE

distinguish Because of the reason mentioned above in


addition to the classical networls with error baclsropagaton
a non-typical tool, ie. ART trpe neural network, has been
used for completion of the described task as it exhibits a
unique ability to determine spontaneously the nnmber and
nature ofthe sledficequivalence classes for the signal input
into the network
In the course ofthe shdy the effect of the following elements

hasbeendyed:
thelearningset

structureoftheneuratnetw&
selection of the learning coefficients
m the criteria for termioation of the learning proms,
affeaingthepocessofneuralnetworkslearning.
certainrules have been elaborated for the considered tasks,
concerning the selection of the learning set, selection of the
network s!mchm,and the training procffs. It turned out that
specific conditions cm be pointed out, common for a l l tasks
concerning the evaldon and c w c a t i o n of the deformed
speech signal. hrdag the research uiteria for temhtion of
the learning prooess has been dekmuwd Normalized
parameter, specifying the level of class recognition
reliabii, e.g. the DELTA factor has been also determind:
DELTA = 1-p s ( Y )- neg(Y)
where:
=

'

POS(Y) = max Zi(l - Y


tiiio

Conclnsions
The obtained &ts
confirm the assumjXion, that the n e d
networks technique can be a useful tool for evaldon of
pathological speech. The ultimate goal of the research is a
construction of a diagnostic system for wide range of
pathological speech varieties.
Bibliography
1. TadeusiewiclR, Wszdek W., Izwomki A.: Application
of Neural Network in Diagnos's of PaIhological Spech,
In: NC'98, Mernational ICSUIFAC Symposium on
N e d Conq~tation,Vtenna, Austria, September 23-25,
1998
2. Tadeusiewia R, Jzwomki A, Wszolek W.:
Pathological Speech Evaluation Using the Artijcial
InMigence Methods, In: World Congress on Medical
Physics and Biomedical Egin&g
September 14-19,
1997, Nice, Franoe

927

Вам также может понравиться