101116_BK_Telekom-CustomerService_DFKI
Transcrição
101116_BK_Telekom-CustomerService_DFKI
Innovative Speech Technology in Customer Service. Dr. Bernhard Kaspar Deutsche Telekom Laboratories, November 2010 Life is for sharing. Agenda. T-Labs and Customer Service Speech Innovation – Examples Beyond Speech Innovation Challenges Conclusion Deutsche Telekom Laboratories 2 Deutsche Telekom Laboratories. Deutsche Telekom Laboratories 3 Telekom Laboratories Telekom and TU Berlin join up for the best of two worlds. Strategic Research Laboratory Interdisciplinary project teams Innovation Development Laboratory Project 1 Project 2 Project 3 Deutsche Telekom Laboratories 44 Telekom Laboratories Open Innovation with institutes across the world. Rheinische Friedrich-WilhelmsUniversität Bonn Imperial College London Columbia University Boston University University of Illinois École Nationale d’Ingénieurs de Brest Univeridad Carlos III de Madrid Technische Universität Darmstadt Stanford University Deutsches Forschungszentrum für Künstliche Intelligenz École Polytechnique Fédérale de Lausanne Norwegian University of Science and Technology Technische Universität Berlin Fraunhofer-Institut für Nachrichtentechnik HeinrichHertz-Institut Fraunhofer-Institut für Offene Kommunikationssysteme Ben-Gurion University Ludwig-Maximilian-Universität München Technische Universität München Universität St. Gallen Deutsche Telekom Laboratories 55 T-Labs have supported Telekom Deutschland to unify voice platforms in customer service. Earlier Today = Telekom Deutschland Festnetz Kunden 26 Mio Breitband Kunden 14 Mio Mobilfunk Kunden 37 Mio 30 Mio calls/month Largest voice portal in Europe Deutsche Telekom Laboratories 6 One voice platform - Synopsis. State-of-the-art technology for customer service. Features Benefits State-of-the-art dialog design and prompting Open dialog and natural language understanding (GPS grammars) Voice biometry capabilities Telekom TTS (scheduled for 2011) Static menus Single words Automated skill based routing Entire self service of selected service cases E.g. line check, status info .. Flow based dialog Domain grammars Deutsche Telekom Laboratories 7 One voice platform - Overall goal Automate standard cases – save ressources for difficult problems. Automate standard cases to save ressources for the hard problems Guarantee quality Agent Ressources 100% Contacts Repeated calls avoid Standard concerns automate Valuable contacts intensify Agent busy time This is a continuing process Deutsche Telekom Laboratories 8 Agenda. T-Labs and Customer Service Speech Innovation – Examples Beyond Speech Innovation Challenges Conclusion Deutsche Telekom Laboratories 9 Innovation …. .. has to be implemented without disturbing the customer. Dear customer, our voice portal is still in a training phase. It will understand you much better until next year. Thank you for your patience. What might be common in other areas .. .. is not accepted in customer service. Deutsche Telekom Laboratories 10 Innovation. .. has to meet the real problems. A brilliant concept .. .. does not always fit to the real life problem. Conditions for innovation New technology Has to be integrated into existing platforms Has to fit into existing/planned workflows Deutsche Telekom Laboratories 11 Innovation Examples. Speech-to-text may help to avoid waiting loops …. Scenario VoiceBox/Callback Currently, all operators are busy. You may leave your concern on a box and we call you back. Customer Leaves message describing concern on VoiceBox I have problems with my bill … REC Speech-to-Text I have problems with I bill You called us an hour ago. I have checked your bill … System Generates service ticket Turns message into text document Agent Opens ticket, when crunch time is over Prepares callback Calls back customer Deutsche Telekom Laboratories 12 Innovation Examples. …. but Speech-to-text is not enough! Es sind mehrere Sachen ich hab ihnen vor zwei bis drei Wochen circa einen Brief geschickt und noch keine Antwort ähm also ich hatte eine Tarif also da ich nicht wusste das es für diese drei Monate auch was kostet ähm bitte stornieren danke auch. Text analysis Metadata Rückrufnummer: 06151-937-9999 Kategorie(n): Rechnung Text Es sind mehrere Sachen ich hab ihnen vor zwei bis drei Wochen circa einen Brief geschickt und noch keine Antwort ähm also ich hatte eine Tarif also da ich nicht wusste, das es für diese drei Monate auch was kostet ähm also bitte stornieren danke auch. Deutsche Telekom Laboratories Processing Steps Goal Agent should be able, to understand concern quickly Steps Speech-to-Text Text formatting Categorization Highlighting of key phrases Adding of metadata (e.g. phone # 13 Innovation examples. Other fields for Speech-to-Text Typical Speech-to-Text Applications In customer service Analyse customer interviews … ich musste viel zu lange warten . Analyse sessions for agent coaching or process insights In communication services „Teasering“ of voice mails Voice-Mail-to-Text 1010010101000111010 Hi Joe, I tried to reach my 0110101010010110100 When can we meet us? I 1001010101010101010 How do you do? Why do 1010101010101011101 we wait on our customer I 0101010000001010101 love you ver much and so 0111010101010100101 this is my book and there 1010111010110010110 you can see our bedroom 1010100000010101010 where we sleep together 0101010101010100101 every day! Could I send 0101010101010111010 you a message about the Deutsche Telekom Laboratories 14 Innovation examples. Classification helps to adapt dialogs. Emotion Classification Speaker Classification Age Seniors V (65+) 65 female male IV (50-65) 50 III (35-50) 35 Dialog adaption II (20-35) 20 I (-20) kids Adaption options in IVR dialog Adapt offers according to age/gender Prioritize waiting list … Deutsche Telekom Laboratories 15 Innovation examples. Classification helps to adapt dialogs. Emotion Classification Speaker Classification Age Seniors V (65+) 65 female male IV (50-65) 50 III (35-50) 35 Dialog adaption II (20-35) 20 I (-20) kids Adaption options in IVR dialog Adapt offers according to age/gender Prioritize waiting list … Deutsche Telekom Laboratories 16 Innovation examples Language detection enables multi-lingual service. Multi-lingual prompting „Cocktailparty-Effect“ as a starting point in dialog Language adaption in dialog Language can be switched after few customer utterances Standard greeting Cocktailparty greeting Language identification Language is identified via both speech recognition an standard language ID methods Deutsche Telekom Laboratories 17 Agenda. T-Labs and Customer Service Speech Innovation – Examples Beyond Speech Innovation Challenges Conclusion Deutsche Telekom Laboratories 18 Customer Service. A case for multi-modality. … you offer a mobile – how does it look like? Voice only just follow this link: http://www.tmobile.de/topangebote/0,17610,1765 1-_,00.html?WT.srch=1 A case for multi-modality Speech is the most natural form of interaction But sometimes, audiovisual interaction is Easier More efficient Audio-visual look here Deutsche Telekom Laboratories 19 Customer Service. Customers want to be served over many channels. From hotlines to multi-access Today Tomorrow Still the top access: The phone Growing rapidly: Mobile Apps The Web alternative: Telekom Service Site lost automated Agent IVR Mobile Apps Web others Deutsche Telekom Laboratories Social media for service: Telekom Twitter TV as service channel? Service via Entertain 20 Customer Service. Case mobile Apps – Prototype „MyAssistant“. Prototype idea A mobile App as a central entry to Telekom service Takes care of orders, devices, tariffs, acount, … So far, without speech technology, but ….. Deutsche Telekom Laboratories 21 Customer Service. Concepts of multi-modality to be migrated to Apps Mobile solutions demand for multi-modal interaction, with speech as an integral part Speech input for mobile devices is (so far) better accepted than voice portals Concept borrowed from another prototype Acoustic Signaling Feedback Vibration Feedback Tilt up Moves focus to upper element Slide Scrolls list Push-to-talk button Tilt left Action as shown on left side bar Voice commands Interpreted by embedded ASR Tilt right Action as shown on right side bar Tilt down Moves focus to lower element Deutsche Telekom Laboratories 22 Customer Service. Case Twitter Telekom hilft Status Small sized solution with surprising success Agent based Ideas for growth and automation Scan and analyse tweeds Select the relevant ones for reaction „Skill based routing“ to aent groups or experts Not speech, but language technology Deutsche Telekom Laboratories 23 Agenda. T-Labs and Customer Service Speech Innovation – Examples Beyond Speech Innovation Challenges Conclusion Deutsche Telekom Laboratories 24 Customer Service and Speech/Language Technology. Selection of remaining challenges (1/2). Make model tuning faster (and cheaper) Support the manual process by automization Update SLM Florian Metze Inspect & tune Run in service Transcribe utterances Detect out-of-vocabulary words Scan STT transcriptions Detect suspicious phonetic strings Cross-check with lexika and alert. Tanja Schulz Hab ihr auch Ei tun sim Angebot Ich gelesen von seit juni Zeitung Deutsche Telekom Laboratories 25 Customer Service and Speech/Language Technology. Selection of remaining challenges (2/2). Speech recognition: How to marry grammars and SLMs Statistical models are not optimal for some cases (e.g. dates) Can grammar rules be included in SLMs without blowing them up? Speech recognition: Beyond word error rates Find better criteria for tuning of speech/language models Sebastian Möller Motivation on next slide Towards practical rules for multi-modal interaction A multi-modal analogon to dialog (speech) act theory? EU project ExtraLing? Deutsche Telekom Laboratories 26 ICS Phase III – Speech Technology and Applications. Performance Criteria – Motivation. STT output example … dein Anliegen überhaupt nicht bearbeitet worden ich habe mir eine Änderung der Tarifoption obwohl man ständig Kundenberaterin fragte hat sie hier nicht und anschließend ich aus der Leitung geschmissen worden sind die wunderschönen Befragung sein getrennt … Sich bin nicht weitergekommen mit meinen Anliegen Herrn Dobermann in 13113 Bonn [eine] Ich habe keine Beschwerde Most likely, awful German Most likely, high word error rate Anyway, main issue quickly understood by agent Beyond word error rates What can be tolerated? We need empirically validated evidence of what is tolerated Deutsche Telekom Laboratories The main issue is not, to optimize word accuracy Rather, we should optimize the „comprehension rate“ Criteria might change according to context. Goal Define „semantics based“ optimization criteria for model tuning. 27 Agenda. T-Labs and Customer Service Speech Innovation – Examples Beyond Speech Innovation Challenges Conclusion Deutsche Telekom Laboratories 28 Customer Service. Perspectives. Future Customer Service Multi-channel Personalized Help customers to help themselves and each other Automated, where possible Via agent, whenever necessary Innovation needed, but More than speech More than technology Deutsche Telekom Laboratories 29 NN. NN Innovation is not always apparent Deutsche Telekom Laboratories 30 Thank you. Life is for sharing.