Download - Natural Voice Recognition
Thomas KrippgansEmail: [email protected].: + 49 731 3994 106FAX: +49 731 3994 251
Natural Voice Recognition
May 2000 Krippgans
R & DR & D
Speech ProcessingSpeech Processing
TelecommunicationTelecommunicationAutomotive
TEMIC5.300 Employee
$ 800 Mio. turn over
Embedded
May 2000 Krippgans
Locations, Employees, Capabilities in Speech Processing
Auburn Hills: 1 employee Sales10 Key Account
Ulm-TEMIC: 75 employeesAcousticsVoice RecognitionDialog DesignIntegration
Ulm-DC RC: 35 employeesAcousticsRecognition (NLU)SynthesisVerificationText Interpretation
Bangalore: 4 employeesautm. Transcritption
Palo Alto: 40 employeesTelematicsCommunication systemsSpeech RecognitionMobile Internet
May 2000 Krippgans
TelecommunicationTelecommunication AutomotiveAutomotive Embedded SystemsEmbedded Systems
DBDBDeutsche Deutsche BahnBahn
ToshibToshibaa
In 1999 about 12.000 PortsIn 1999 about 75.000 Units
ThomsonThomsonmultimediamultimedia
Launching Customer
Belinguasoft-CAD Systems
THB Bury
Tobit
At the beginning is:
““Dada”Dada”
May 2000 Krippgans
Natural Voice
Dadaiiiiiii
May 2000 Krippgans
Natural Voice
Papa 8-)8-)
8-)8-)
8-)8-)
May 2000 Krippgans
• In age of 7 to 10 month kids start to In age of 7 to 10 month kids start to move their lower jawmove their lower jaw
•every of them, in over more than 27 every of them, in over more than 27 different languages, usedifferent languages, use
•““Dada” “Mama” “Gogo” as the Dada” “Mama” “Gogo” as the common wordscommon words
• this kids use the so called Protowords this kids use the so called Protowords will be find in all languages will be find in all languages
Natural Voice
May 2000 Krippgans
Natural Voice? ?? ? ! !! !
????To be or not
to be
that’s the question !?
May 2000 Krippgans
For applications in the world of For applications in the world of Service Provider using Natural Service Provider using Natural Language Recognition on thing is Language Recognition on thing is importand:importand:
•Transaction Success Rate (TSR)Transaction Success Rate (TSR)
Natural Voice Recognition
May 2000 Krippgans
... and the system picks key words (word spotting).“I would like to record a message”
... and the system picks key phrases (phrase spotting).“Tomorrow I would like to go from Ulm to Munich ”
What is Natural Language Understanding?Some definitions:
The user can say anything she/he wants...
“Do I have a new message?” vs. “I would like to record a new message”
... and the system recognizes all words and attempts to understand them. (Word Hypothesis Graph and Parser from Temic )
May 2000 Krippgans
speech signal
recognition result
parsing result
Natural Language Understanding
May 2000 Krippgans
Results from a Field Trial
• Support Hotline System; Experience Support Hotline System; Experience from the ACCeSS Project (EU founded)from the ACCeSS Project (EU founded)
• A incoming call routing systemA incoming call routing system
• Natural Language Recognition 2nd Natural Language Recognition 2nd Generation (Parsing and NLU Generation (Parsing and NLU Dialogmanager)Dialogmanager)
• Evaluation of 1.500 Dialogues during a Evaluation of 1.500 Dialogues during a three months field trialthree months field trial
• Installed in a Call Center enviromentInstalled in a Call Center enviroment
May 2000 Krippgans
•We evaluated 1,528 dialogues with We evaluated 1,528 dialogues with
9,159 recorded utterances, 12,886 9,159 recorded utterances, 12,886
total wordstotal words
•Dialogue Duration 70 secDialogue Duration 70 sec
•Hang-Ups 13 %Hang-Ups 13 %
•Average Success Rate 97 %Average Success Rate 97 %
Results from a Field Trial
May 2000 Krippgans
Spoken Language DialogueMain components of spoken language dialogue systems
recognition
acou
sti
c d
ata
AS
CII
understanding
info
rmati
on
, m
ean
ing
dialogue planning
next
dia
log
ue s
tep
May 2000 Krippgans
ACD
Database server
Operator
LineInterface
SystemControl
SpeechSynthesis
DialogueManager
DatabaseInterface
SpeechRecognizer
LAN
User
Call Center Integration
May 2000 Krippgans
How many Users are Out There?
May 2000 Krippgans
Solutions for huge subscriber bases
StarRec KXL (PCI/cPCI)
16 ASR Ports ==
240 ASR Ports 3030ServerServer
==
2.400 ASR Ports ==
May 2000 Krippgans
Speech Recognition Solution
(High Integrated NLU)
- 19 inch slide- in module- 19 inch slide- in module
- up to 240 ports NLU;- up to 240 ports NLU;
good for 600 telephone good for 600 telephone
portsports
- comfortable maintenance- comfortable maintenance
- a way to design huge - a way to design huge
systemssystems
May 2000 Krippgans
Tool‘s GDS (Grammar Design Software)
May 2000 Krippgans
No Transaction Success Rate
Dadaiiiiiii ??