A system and method for authoring a voice application is provided. The voice recognition process uses a dynamic grammar which obtains data from a backend data source based upon an input in order to create expected results that a speech engine can recognize. The process can retrieve data from at least one of a plurality of back-end data sources and can build a grammar based on the data using at least in part a dynamic grammar builder. The grammar is loaded into the voice recognition application using at least in part the reusable dialog component. A data access service, XSLT processor, or other data accessing framework can be used to facilitate access and manipulation of data in heterogeneous environments.