A Unified Language Model Architecture for Web-based Speech Recognition Grammars

Supporting popular language model grammar formats, such as JSGF and XML-SRGS, has been an important step forward for the speech recognition community, particularly with respect to integration of human language technology with Internet-based technologies. Industry standard formats, though conceptually straightforward implementations of context free grammars, contain restrictions that pose serious challenges when applied to aspects of the speech recognition problem. This paper compares and contrasts these formats, discusses the implications for speech recognition systems, and presents a unified language model architecture to support transparent conversion between various language model formats. This architecture requires the conversion of higher-level grammar specifications such as the JSGF and SRGS into lower-level theoretical structures such as Augmented Backus-Naur form and Standard Backus-Naur form. The public domain implementation of this architecture provides a framework for future advancements in language model conversion and Web-based speech recognition applications