Hierarchical search for large-vocabulary conversational speech recognition - Working toward a solution to the decoding problem

The aim of a continuous speech-recognition system is to provide an efficient and accurate mechanism to transcribe human speech into text. To make this system ubiquitous, it is important that the system must be able to handle a large vocabulary, and be independent of speaker and language characteristics such as accents, speaking styles, dysfluencies, syntax, and grammar. The problems of search for large vocabulary continuous speech recognition (LVCSR) systems are introduced. A typical implementation of a search engine is discussed and the efficacy if this approach on a range of problems is demonstrated.