120 likes | 255 Views
Speech Recognition and Machine Translation. Stephan Kanthak s.kanthak@aixplain.de AIXPLAIN AG, Aachen, Germany. Technology. Focus on Statistical Methods Modular Software Architecture: Speech Recognition Machine Translation Retrieval Combined Applications: Multilingual Retrieval
E N D
Speech RecognitionandMachine Translation Stephan Kanthak s.kanthak@aixplain.de AIXPLAIN AG, Aachen, Germany
Technology • Focus on Statistical Methods • Modular Software Architecture: • Speech Recognition • Machine Translation • Retrieval • Combined Applications: • Multilingual Retrieval • Retrieval of Multimedia Data • Speech-To-Speech Translation • ... • Original Mission: • Speech-To-Speech Translation (for Mobile Devices) Speech Recognition Machine Translation Knowledge Management
Speech Recognition Preprocessing Search Acoustic Models Pronunciation Lexicon Language Model Result
Speech Recognition: Current Research • Feature Extraction • Phase Information • Noise Robustness • Acoustic Models • Adaptation • Maximum Entropy Models • Discriminative Training • Multilingual Acoustic Models • Language Models • Class-Based Backing-Off Models • Open Vocabulary • Search • Weighted Finite-State Transducer • Dialogue Management • Reinforcement Learning
MT: Alignment Templates Preprocessing Search Lexical Models Alignment Templates Language Model Postprocessing
Machine Translation: Current Research • General • Incorporate Syntactic Knowledge • Improved Alignment Models • Context-Free Grammars • Phrase-Based Models • Language Models • Class-Based Backing-Off Models • Maximum Entropy Models • Search • Stochastic Parser • Weighted Finite-State Transducer
Research Projects • At Lehrstuhl für Informatik VI • Current Projects: • PF-Star: Speech Translation • TransType2: Computer Assisted Translation • LC-Star: Lexica and Speech Corpora for Speech-To-Speech Translation • Completed Projects: • TC-Star_P: Preparation of TC-Star • CORETEX: Core Speech Recognition Technology • VERBMOBIL II: Speech-To-Speech Translation • Eutrans: Speech-To-Speech Translation • ADVISOR: Transcription of German Broadcast News and Retrieval of Videoclips • GIZA++: Open-Source Statistical Machine Translation
Success Stories • Research Systems: • DARPA HUB4 Evaluation‘99: 4th best ASR system • VERBMOBIL: fastest and 2nd best ASR system • VERBMOBIL: best MT system • DARPA MT Evaluation‘02: best MT system • DARPA MT Evaluation‘03: 2nd best MT system • Production Systems: • Almost 200 speech recognition licenses sold in 2003 • Largest installation of a pick-by-voice system in Europe in December 2003