Speech Projects at ITC irst Roberto Gretter ITC
Speech Projects at ITC irst Roberto Gretter ITC irst AT&T Shannon Labs, Florham Park, NJ, 6/27/2000
ITCirst Trento, Italy § About 100 researchers in 5 divisions § division SSI: vision, speech, statistics, software system analysis § speech: 16 researchers in 3 internal projects: l l l MUNST - multilinguality, broadcast news, CORETEX SHINE - recognition in car, acoustic localization DITELO - telephone, dialogue
Accessing tourism information § Local tourism agency provides info database about accomodation, structures and services, localities, events, sports, leisure time, art, nature, transportation, … (see http: //www. provincia. tn. it/apt go) § Mixed initiative dialogue, explicit confirmation, no barge -in, over 2000 words vocabulary, information presented using a Mixed Representation approach (template-based + deep generation) § First prototype for data collection l Bootstrap system handles accommodation and place names l Bootstrap grammars
Tourism information § Bootstrap grammars: <localita`> <vorrei> trento vorrei rovereto dovrei riva del garda voglio cles mi piacerebbe <stelle> ( a | con) ? (( 1|2|3|4|5 ) o ? )+ stelle <tipo> albergo campeggio residence affittacamere <vorrei> andare in un <tipo>/TYPE <stelle>/STARS <vorrei> soggiornare vicino a <localita`> /CITY un <tipo>/TYPE a <localita`> /CITY <stelle>/STARS mi manda i dati <output>/CHAN al numero <tel>/FAX <vorrei> il <info>/INFO del <tipo>/TYPE <n>/NAME § Only 2 main grammars defined § Data collected without Wizard of Oz main_grammar
Tourism information: acquisitions § Users provided with some task to accomplish. § Users: domain expert / non expert § Acquisitions using the first prototype (May 2000): 2 hours 37 min speech - about 220 dialogues § Preliminary evaluation on 34 dialogues
Aurora (proposal) § Similar to HMIHY task: l l call routing (How can I help you? ) natural language processing, information extraction, document analysis and retrieval, database access automatic routing agent automatic answering agent user human router l German (DFKI), Italian (Offnet, IRST, CELI), Dutch (CTIT, Uni Twente), Spanish (DFE, Uni Barcelona) § Multilinguality in Spee. Data l human operator data-entry (Land-register) test-it test-de HMM-it 93. 9 HMM-de 89. 4 HMM-mix 94. 3 89. 5
Kataweb § One of the most important Italian WEB portals go • Requirement: Fast Porting of new speech and NL technologies to WEB portals (Dec 2000) l Access information via written queries: • Who invented the electrical light? • Which is the capital of Alaska? l Information extraction from news • The Bank of Japan decided … the president said. . . l Spoken dialogue by phone • call center solution • trading on line by phone, about 500 stocks § Future developments are possible l l l extend domains integrate speech and natural language computer vision
Technology transfer to solution developers § What they are asking us, today l l to include the recognizer in menu-based call centers to build applications requiring dialogue § What they are going to ask us, tomorrow: l l multilinguality speaker verification § What we provide l l Spinet server (recognizer, grammars, lexicon, . . . ) API (c++, Java) assistance in building prototypes/systems assistance in verifying/improving systems
- Slides: 8