Determination of optimal training methodologies for discrete/dependent speech recognition (SR) systems

40 Pages·0.44 MB·en_US

by  Rhoads, Mark Charles.;Poock, G. K.(Gary K.)

*AVA ADLMT£
,
MOWTfcREy. CALIFORNIA
93943-5002

-•
NAVAL POSTGRADUATE SCHOOL
Monterey California
,
THESIS
Determination of Optimal Training
Methodologies for Discrete/Dependent Speech
Recognition (SR) Systems
by
Mark C. Rhoads
March 1992
ThesisAdvisor: Garv K. Poock
Approved for public release; distribution is unlimited.
258482
T

)
Unclassified
SECURITY CLASSIFICATION OFTHIS PAGE
REPORT DOCUMENTATION PAGE
ia report 5EOTHTT CLASSIFICATION1 ib RESTRICTIVE MARKINGS
Unclassified
2a" SECURITY CLASSIFICATION AtTHORrTY "5! DISTRIBUTION AVAILABILITY OF REPORT
Approvedforpublicrelease;distributionisunlimited
&. ^CLASSIFICATION doWngRAdIng schedule
4. PERFORMING ORGANIZATION REPORT NUMBERS) J. MONITORLNG ORGANIZATION REPORT NUMBER(S)
g NAME 6F BRFftRMNfl 6RgaNBATH6n' A. WFKISYMB&L 7,. NAME6TMontToRNg" 5RganTzATIon"
NavalPostgraduateSchool (IfApplicable NavalPostgraduateSchool
6c ADDRESS(city, stale, andZIPcode) "%. ADDRESS(city, state, andZIPcode)
8a. MNoAnMtEerOeFy,FCUNADI9N3G94S3PO5N0S0O0RING A bFHcISYMboL TMKoonclTejr;7rReeE\yM,\ECc)5A3T9N3s9T4R3u-5M0E0x0T EETTriFKKTRSR NUMBER
ORGANIZATION (IfApplicable)
8c ADDRESS(coy, ttc/r a/ufZIPcode) 10 SOURCEOF FUNDING nTJKJBEIS
.
pRckSRaM PROJECT TXsTT W6RK IWT
element no NO. NO. ACCESSION NO.
11 I'liLF (IncludeSecurity Classificatum)
DeterminationofOptimalTrainingMethodologiesforDiscrete DependentSpeech Recognition(SR) Systems (Unclassified)
15 Personal AuTH6R(S,
LCDR Mark C. Rhoads
13a TYpEbFREPoRT 73T TIME COVERED 77 DATEOFREPORT(year. month.doM n pagec^t
Master'sTVhesis FROM OCT 90 TO M\R 92 March1992 30
16 sTTPLEMT rARY SfoTATV>K'
Theviewsexpressedinthis thesisare those ofthe authoranddonotreflecttheofficial policy orpositionofthe Department of
Defenseorthe IS. Government
[T TosAtt Wbte 18 SUBJLCl TERMS(continueon reverse Jnecessary andidentify by block number)
TteTd" GROUP SITJgroI'p Speechrecognition;trainingmethodologies,experimentalresults;conclusions
19 ABSTRACT {Continueon reverse ifnecessary andidentify by block numberi
Aresearchexperimentwasconductedtodeterminewhethervariouscombinationsoftrainingmethodologiesandspeaking
voiceswouldaffectrecognitionaccuraciesamongstumquespeakerdependentSpeechRecognition(SR)systems. Theexperiment
usedaSRsystem(VOTAN VTR605011)whichisbasedonVOTAN (proprietary)technology. Tensubjectstrainedfivedifferent
voicepatternseachandconductedfournatural voiceteststocompilestatisticsabouttherecognitionaccuracyforeachpattern.
Twopatterns(naturalvoiceanddeclarativevoice)wereretestedusingadeclarativevoice.
Theexperimentwassuccessfulanddemonstratedthatdifferentcombinationsoftrainingmethodologiesandspeakingvoices
cansignificandyaffecttheperformanceofuniquediscretedependentSRsystems. Thisthesisdiscussestheresearch
methodology,reviewsandanalyzesthedatacollected,andstatesconclusionsdrawnabouttheparticulardependentSRsystemused
intheexperiment
io Dabibvlkm'AvukbilitvofAbstract n ii absTRacT SECURITY clA^IficaTIon"
§XIdwaficd/wfanilcd Dsame *- rpt DT1C users Unclassified
NaMe 6+ responsible EJdTvECaT 1Tb TELEPHONE (Include Area 12c OFFICESYMBOL
Gary K Poock Code) CodeORPK
408-646-2636
DOFORM 1473 84 MAK 83APRedition nw> be used untilexhausrd SECURITY CLASSIFICATION OFTHIS PAGE
AJIothereditionsare obsolete Unclassified
Approved for public release; distribution is unlimited.
Determination of Optimal Training Methodologies
for Discrete/Dependent Speech Recognition (SR) Systems
by
Mark C. Rhoads
Lieutenant Commander. United States Navy
B.S.. University ofKansas. 1978
Submitted in partial fulfillment ofthe requirements for
the degree of
MASTER OFSCIENCEININFORMATION SYSTEMS
from the
NAVAL POSTGRADUATE SCHOOL
March 1992
ABSTRACT
A research experiment was conducted to determine whether various
combinations of training methodologies and speaking voices would affect
recognition accuracies amongst unique speaker dependent speech recognition
(SR) systems. The experiment used a SR system (VOTAN VTR 6050II) which is
based on VOTAN (proprietary) technology. Ten subjects trained five different
voice patterns each and conducted four natural voice tests to compile statistics
about the recognition accuracy for each pattern. Two patterns (natural voice and
declarative voice) were retested using a declarative voice.
The experiment was successful and demonstrated that different
combinations of training methodologies and speaking voices can significantly
affect the performance of unique discrete dependent SR systems. This thesis
discusses the research methodology, reviews and analyzes the data collected, and
states conclusions drawn about the particular dependent SR system used in the
experiment.
in
A3733
(L.I
TABLE OF CONTENTS
INTRODUCTION
I. 1
BACKGROUND
A.
1
PROBLEM
B. 2
SCOPEOFTHETHESIS
C. 2
D. LIMITATIONS 3
EXPERIMENT PROCEDURE
II. 4
A. SUBJECTS 4
B. SR SYSTEM 4
C. EXPERIMENT DESIGN 5
PROCEDURE
D. 6
1. Training 6
2. Testing 7
'r
E. INDEPENDENTAND DEPENDENTVARIABLES 8
III. RESULTS 9
A. OVERVIEW 9
1. Analysis ofVariance 9
2. Impact of Variables 10
a Subject' Variable 10
b. "Trial' Variable 10
c. 'Pattern' Variable 12
d. 'Voice' Variable 12
B. DISCUSSION 17
CONCLUSIONS
IV. 19
REFERENCES
21
APPENDIX A 22
INITIAL DISTRIBUTION LIST 23
IV