Methodologies for language modeling and search in continuous speech recognition

Abstract

Automatic speech recognition has made significant strides from the days of recognizing isolated words. State-of-the-art systems are capable of recognizing tens of thousands of words in complex domains such as newspaper correspondence and travel planning. A major part of this success is due to advances in language modeling and search techniques that support efficient, sub-optimal decoding over large search spaces. The benefit from focusing a recognition system on a particular domain has motivated a steady progression from static language models towards more adaptive models that consist of mixtures of bigrams, trigrams and long-distance n-grams. Similarly, the availability of multiple sources of information about the correct word hypothesis has led to the advent of efficient multi-pass search strategies. The result is a powerful pattern-matching paradigm that has applications to a wide range of signal detection problems. Future research in large vocabulary continuous speech recognition will be directed towards developing more efficient means of dynamically integrating such information.