Network Language Model Generated from a Modification Structure-based Language Model Template 2009-01-1478
Natural speech contains a wide variety of expressions, making it much more difficult to recognize than specified verbal commands. We have developed a method using a language model template to generate language models capable of recognizing manifold expressions. This template method is designed on the basis of the modification structure of keywords. The evaluation results for a language model generated with this method showed that it allowed 630,000 command patterns with a high recognition rate of 96%. The results also revealed that the processing speed and memory used for speech recognition with this language model were at a practical level.