Sogang University: Dept of Computer Science 구명완교수 서강대학교컴퓨터공학과 Email: mwkoo9@gmail.com
내용 개념도 시맨틱해석기 대화상태추적기 대화관리기 자연언어생성기 결론
개념도 음성인식 음성대화
음성대화솔류션사례 음성대화서비스 음성대화로봇 Apple Siri Google MS Cortana Amazon Echo 삼성전자 Softbank Nao/Pepper MIT Jibo 형상 상황인지기 특징 Chat 개념첫도입 26종류의목적지향 홈킷과연동 반대화 ( Google Now) Gmail, 위치정보및구글검색활용 Google Assistance + Google Windows 10 PC & phone 기반 Google now, Siri 기술결합 8 종류의목적 지향 아마존서비 스특화 영어만지원 홈넷과결합 S보이스 비브랩스 (ViV Labs) 인수 S8 에탑재예정 대화 + 홈넷 + 서비스 ( 호텔등 ) 감정인식 비언어적메시지인식 클라우드기반학습 사용자선호도학습 미출시 Home
Siri vs Google Now vs Cortana Comparison: https://www.youtube.com/watch?v=1ibu7gz-jh8
Siri vs Google Assistant Comparison: https://www.youtube.com/watch?v=jfiu5rfnhzo
개념도 The task of a dialogue system to understand what the user wants at any point in a conversation Problem
개념도 A dialogue system Ex: System 2: You are looking for the train station? The driver replies No. The only remaining valid hypothesis, petrol station
Goal Basic Concept for Spoken Language System One day humans will be able to speak to computers as human do among humans a spoken dialogue system A system that can maintain an intelligent conversation with a human, via speech, and in real time, both input and output is speech dialogue is represented a sequence of dialogue turns each turn : a system utterance followed by a user utterance
Basic Concept for Spoken Language System
Basic Concept for Spoken Language System A statistical spoken dialogue system to build a data-driven dialogue system improves over time, behave in human-like way Speech recognizer converts the user s speech into text produces every possible hypotheses of what the user said Open Source : Kaldi Semantic decoder extracts a semantic representation of the user s intention, represented by semantic units called concepts produces all possible semantic concepts Machine Learning, DNN
Basic Concept for Spoken Language System Dialogue management + Dialogue State Tracking selects the most appropriate system output in semantic units to make use of the alternative options in the input particularly useful in noisy environments, DNN + Reinforcement Learning(Deep RL) Natural language generator transforms the semantic notation back into a text representation Machine Learning, DNN Speech synthesizer converts this text into speech Open Source: Festival
Basic Concept for Spoken Language System
Example dialogue
Example dialogue
음성대화 인터페이스 What is CookTalk? A prototype for spoken dialogue system: Page 16
LG DIOS Steam & Microwave Oven Some Issues of Current Steam & Microwave Oven So many menus ( each menu has unique menu number ) Don t know menu name exactly May need search by ingredients, by accessories, by mode inform ingredients for each menu Various accessories depending on cooking App. which control Oven by NFC Add spoken dialogue interface to this App. Easy steps for cooking Page 17
LG DIOS Steam Oven & Microwave Oven Oven Accessories Metal Tray( 법랑접시 ), Beetle Shelf( 석쇠 =grate), Grill Pan( 구이전용팬 ) Page 18
Characteristics Spoken Language Dialog Interface Search by menu Search by ingredients(material) Search by mode(type) Search by accessary(con) Provide necessary ingredient list Provide guide 무엇을요리하시겠습니까? How may I help you 치즈케이크 Cheese Cake 가지고있는재료를말씀해주세요. You may know the menu by ingredient Potato Ultimately, user can give orders using spoken dialogue interface Page 19
CookTalk Real Challenge 2015
: Future 기술 : DNN + Reinforcement Learning End-to-End Network 서비스 목적지향 + 쳇 Multi-task 학제간융합교육필요 공개오픈소스를활용한다양한프로젝트가능