An efficient vocabulary reduction algorithm for a very large vocabulary continuous speech recognition engine

The need for automatic speech recognition engines that can handle very large vocabularies and are useful for a wide range of applications has become imperative in the last decade. Applications using spontaneous speech recognition use extremely large lexicons that lead to high computational complexity when traditional approaches are used. The research presented here proposes a design for a speech recognition engine that significantly improves computational complexity by decreasing the size of the active vocabulary used in the recognition process. The engine is based on a Multi-Stage structure that splits the process into three stages. The results show that, despite the implementation of two additional stages, this configuration effectively reduces the overall computational complexity by using smaller vocabulary size and has a minimal affect on speech recognition performance.