Adaptive Language Model in Automatic Online Subtitling

A. Pražák, J.V. Psutka, J. Hoidekr, J. Kanis, L. Müller, and J. Psutka (Czech Republic)


LVCSR, real-time recognition, class-based language model, automatic subtitling


This paper describes an adaptive language model for recog nition tasks, where person names are specific for each recognition session and have to be added to the recogni tion vocabulary. Creation of an adaptive language model with word (name) classes is illustrated on two different tasks of automatic online subtitling: subtitling of parlia ment meetings (1.5 % of names) and ice-hockey matches (15 % of names). The second part outlines the system for automatic online subtitling with vocabulary up to 100 000 words in real-time. The recognition system is based on Hidden Markov Models, lexical trees and bigram language model. Finally, experimental results with and without the adaptive language model are reported and discussed.

Important Links:

Go Back