What makes a successful speech-enabled call routing application? Diana Binnenpoorte and Dorota Iskra LangTech February 28 - 29, 2008 – Rome, Italy 1
Overview • Who are we? • Why speech-enabled call routing? • What is speech-enabled call routing? • Performance measures • Success factors in various phases • Summary LangTech February 28 - 29, 2008 – Rome, Italy 2
Who are we? LogicaCMG and speech technology: • The VOICE team as part of Customer Contact Solutions • Expertise on the design, implementation and integration of telephony self services in existing architectures LangTech February 28 - 29, 2008 – Rome, Italy 3
Examples of applications • CLAIRE : a 24 hour speech-driven operator and receptionist • Mobile ticket service : mixed-initiative dialogue system through which mobile phone users can specify their journey and order tickets • Stock Information Speech Portal : open dialogue system in which users can request share holder’s information, information on funds, tips and so on. The system takes into account the expertise of the user • Call routing applications • Voice Verification : users get access to their accounts based on biometric data of their voice (not yet fully implemented) LangTech February 28 - 29, 2008 – Rome, Italy 4
Why speech-enabled call routing? (1) • Route calling customers to appropriate agent or service in call center • Customers usually press keys: touch-tone IVR • Touch-tone IVR menus remote from customer’s intuition: – Many time-outs – Wrong choices – Transfers to operator – Pressing ‘0’ • IVR menus needed in large organizations: – One contact number – Specialized agents LangTech February 28 - 29, 2008 – Rome, Italy 5
Why speech-enabled call routing? (2) Speech-enabled call routing: • Facilitates same functionality as touch-tone IVR • Open Speech Recognition • Much more customer-friendly: – Customer is free to speak question in a natural way – Customer no longer needs to choose – Shorter and more efficient dialogues LangTech February 28 - 29, 2008 – Rome, Italy 6
What is speech-enabled call routing? un- dialogue decided manager open classifi- speech filtering cation recognition candi- n-best dates list agent of service LangTech February 28 - 29, 2008 – Rome, Italy 7
Performance measures • Success of spoken dialogue systems often expressed in: – Dialogue Success Rate – Recognition accuracy: Word Error Rate • BUT, is the application: – widely used? – appreciated by customers? – appreciated by employees of service provider? • THEREFORE, other factors: – Customer experience – Involvement within organization LangTech February 28 - 29, 2008 – Rome, Italy 8
Various phases of system development Design phase : • Establishing functionalities • Dialogue design • Determining exit points Building phase : • Coding and testing • Collecting training material Deployment phase : • Training call center agents • Business intelligence LangTech February 28 - 29, 2008 – Rome, Italy 9
Design phase (1) Speech Navigation Level 0 and 2 nd ROUND level 1 FROM Identification EXECUTE 1-9-2006 Page 1 Open (from (page 6) 'Voorportaal') Dialogue design Set round + 1 (M2) Collect session data (M1) Round = from Open to Open, excluding DTMF parts Irritation_Q = round dependent irritation factor irritation OPEN Execute NO threshold YES (page 1) (page 6) exceeded? D8 • Contradicting interests: Prompt (pUitlegDemo) Reset round dependent parameter (M3) “Dit is een demo ...” Session data YES DEMO_1? call frequency_novice – Technical performance FLEXIBLE parameter D5 DIALOGUE BLOCK Call Session NO (page 7) frequency_no data ActionLabel parameter vice? D4 NO YES ACTION ROUND = 1? YES – Customer experience LABEL_1? =< frequency NO D1 D2 FLEXIBLE threshold DIALOGUE YES > frequency threshold BLOCK_1? Prompt (pUitleg) D3 Prompt (pActionLabel) Flexible dialogue – Organizational issues parameter “Dit systeem ...” “Actie label” NO (Q1.0a) “Open vraag: waar (Q1.0) “Open vraag: waar belt u nog meer over, anders Irritation_Q+1 belt u over” hang op” Prompt (pNR) • Influences all performance measures TOO_LONG INPUT? NOT_RECORDED “Niet gehoord” D6 Prompt (pBreak, pBreak2) Prompt (pNU) REPEAT <beep> OPEN NOT_UNDERSTOOD “Sorry, onderbreken” “Niet verstaan” (page 2) filtering (M4) Prompt (pNU23) Set location code Access =1 # of EXITS? =0 OR > 3 (M5) table “Niet helemaal verstaan” D7 =2 OR 3 EXIT1 EXIT2-3 (page 5) (page 4) LangTech February 28 - 29, 2008 – Rome, Italy 10
Design phase (2) Determining exit points • Agents are skilled in different tasks, answering various types of questions, i.e. categories • Organization of call center leading in selecting exit points? • Customer centered approach • Number of exit points • Influences all performance measures LangTech February 28 - 29, 2008 – Rome, Italy 11
Building phase (1) Collecting training material , i.e. collection of tagged sentences • “The more the better”? • Well-distributed over exit-points • Training material should resemble real customer questions: – Terminology and formulation – Variance LangTech February 28 - 29, 2008 – Rome, Italy 12
Building phase (2) • Collecting strategies: – Spoken material (live or simulated) – Written material (natural or simulated) • Influences technical measures and involvement within organization LangTech February 28 - 29, 2008 – Rome, Italy 13
Deployment phase Training call center agents • Learn new conversational techniques • Learn how to interpret recognized sentence Business intelligence • Valuable information can be retrieved, e.g. – Statistics, e.g. load effect on call center – Effect of marketing campaign • Both increase the support within organization LangTech February 28 - 29, 2008 – Rome, Italy 14
Summary • Besides commonly used technical performance measures also other factors play important role in the success of a speech-enabled call routing application • These factors are customer experience and involvement within organization • During various phases of implementation these factors play an important role • Sometimes conflicting interests LangTech February 28 - 29, 2008 – Rome, Italy 15
16 Questions? LangTech February 28 - 29, 2008 – Rome, Italy Thank you!
Recommend
More recommend