The automatic synthesis of fluent speech requires robust and accurate natural language processing technology at the word level (word pronunciation) and the text level (prosody generation). Machine learning offers methods to learn these processing tasks and their intermediary prerequisites based on annotated lexicons and corpora.
At the word level we will review machine learning approaches for G2P, tokenization and POS tagging; at the sentence level we focus on phrase chunking and semantic information sources for accent and phrase break placement. We will review the main strands of existing algorithmic solutions (stressing the lazy-eager learning dimension), and describe state-of-the-art sequence processing methods that distinguish between local classification and global search.
Presenters
Antal van den Bosch,
Tilburg University Antal.vdnBosch at uvt.nl
Walter Daelemans,
Univ. of Antwerpen,
Short Bios
Antal van den Bosch is associate professor at the Dept. of Communication and Information at Tilburg University, The Netherlands, heading the ILK Research Group. His work focuses on memory-based machine learning methods applied to NLP (e.g. with Daelemans, "Memory- based language processing", Cambridge University Press, 2005). Van den Bosch currently works on memory-based language models and machine translation.
Walter Daelemans is professor of computational linguistics at the University of Antwerp and director of the CNTS Language Technology Group. He has published widely on Machine Learning applied to Natural Language Processing and NLP applications ranging from speech synthesis to text mining. He is currently associate editor of Research on Language and Computation.
