Statistical language modeling with semantic classes for large vocabulary speech recognition in embedded systems

In this paper we investigate the use of semantic classes in an n-gram language model used for speech dictation in an embedded environment. The alternative to using semantic classes is to use automatically clustered word classes, but with this approach it is often difficult to increase the modeling accuracy and reduce the model size at the same time - two factors that are critical for an embedded application. We describe how the introduction of semantic classes in a large vocabulary (33000 words) embedded dictation task for US English reduced the model size by 16.0%, while at the same time also reducing the word error rate by 12.0% relatively.