Speech recognition and retrieval using unsupervised sub-word language models

Google Tech Talks February, 8 2008 ABSTRACT Unsupervised morpheme analysis and language models developed at Helsinki University of Technology (TKK) open interesting new views on large vocabulary speech recognition, information retrieval and machine translation. This approach learns suitable sub-word units directly from relevant text corpora in a completely data-driven manner and can, thus, be easily ported to various morphologically complex languages. Our system includes several public domain software packages: Morfessor for determining the modeling units, VariKN to train effectively smoothed long-span LMs, a near-realtime single-pass decoder for LMs and LVCSR, and demos at: http://www.cis.hut.fi/projects/morpho/ http://www.cis.hut.fi/projects/speech/ I will also summarize the results from the recent Morpho Challenge machine learning competitions in morpheme analysis for language modeling and information retrieval: http://www.cis.hut.fi/morphochallenge2007/ Speaker: Mikko Kurimo Mikko Kurimo is currently Academy Research Fellow at Helsinki University of Technology (TKK), where he was an acting professor of Computer and Information Science (2001-2003). In 1998-2000 he worked as a postdoc at IDIAP and his Doctor of Science (PhD) degree is from TKK 1997. He has a long research record and a number of publications in large vocabulary speech recognition (since 1990) and has been involved in several international and national research projects and is TKK's site manager in the FP6 Network of Excellence PASCAL. At TKK's Adaptive Informatics Research Centre (which is one of the Centres of Excellence nominated by the Academy of Finland) he leads the Speech Recognition and Multimodal Interfaces research groups. The current research focus of his speech group is in language independent and unsupervised models for continuous speech with morpheme-based language models for very large vocabulary. The relevant pilot applications in the group range from unlimited vocabulary dictation systems for different languages to audio indexing and speech-to-speech translation. At the moment Mikko Kurimo is an International Fellow at SRI until February 2008.