110 likes | 303 Views
1990s DARPA Programmes WSJ and BN. Dapo Durosinmi-Etti Bo Xu Xiaoxiao Zheng. Introduction. 1 Definition of DARPA, WSJ and BN 2 Technology used in the two test beds. 3 Change from WSJ to BN 4 Comparison between WSJ and BN 5 Prospect of ASR 6 Conclusion. 1 Definition of DARPA, WSJ and BN.
E N D
1990s DARPA ProgrammesWSJ and BN Dapo Durosinmi-Etti Bo Xu Xiaoxiao Zheng
Introduction 1 Definition of DARPA, WSJ and BN 2 Technology used in the two test beds. 3 Change from WSJ to BN 4 Comparison between WSJ and BN 5 Prospect of ASR 6 Conclusion
1 Definition of DARPA, WSJ and BN • DARPA- Defense Advanced Research Project Agency • WSJ- Wall Street Journal • BN- Broadcast News
Overview • ASR- Automatic Speech Recognition • In early 1990s-Wall Street Journal • Improvement from Resource Management • 1995-Broadcast news
2 Technology used in WSJ • continuous density HMM with Gaussian mixturefor acoustic modelling • n-gram statistics estimated on newspaper tests for language modelling • bigram and trigram in the graph search strategy • cepstrum-based features, context-dependent phone models, phone-duration models and sex-dependent models.
3 Change occurs from WSJ to BN • WSJ was built in the early 1990s • 1995 the BN test bed was introduced
WSJ Financial domain focus Written language domain Simulated dictation Only speech is used One speaking style and accent One speaker at a time Speaking steadily Find and retrieve word BN National news focus Spoken language domain Real-world, found speech Speech, video and text Every speaking style and accent One or multiple speakers at a time Speak continuous More sophisticated and related search 4 Comparison between WSJ and BN
FUTURE of ASR Spoken language interface applications include voice calling, retrieving and sending email/voicemail; using the internet to program remote speech recognition and collection.
5 Conclusion • WSJ & BN • Technology • Progress
References • “Broadcast News is Good News” Francis Kubala from http://www.nist.gov/speech/publications/darpa99/pdf/sp110.pdf • “Corporate activities in speech recognition and natural language: another ‘new-science’-based technology” Konstantinos Koumpis from http://www.dcs.shef.ac.uk/~costis/pubs/pdf/ijim99.pdf