Lesław Pawlaczyk i Paweł Bosky
Insitute of Informatics, Silesian University of Technology, Akademicka 16, 44-100 Gliwice, Poland
In this article we present a system for clustering and indexing of automatically recognised radio and television news spoken in Polish language. The aim of the system is to quickly navigate and search for information which is not available in standard internet search engines. The system comprises of speech recognition, alignment and indexing module. The recognition part is trained using dozens of hours of transcribed audio and millions of words representing modern Polish language. The training audio and text is then converted into acoustic and language model, where we apply techniques such as Hidden Markov Models and statistical language processing. The audio is decoded and later submitted into indexing engine which extracts summary information about the spoken topic. The system presents a significant potential in many areas such as media monitoring, university lectures indexing, automated telephone centres and security enhancements.
full article: http://www.springerlink.com/content/005r89701h218005/