Methods and apparatus are provided which may be employed to perform speech recognition processing on a grid computing system. In some embodiments, the grid computing system includes a server system which receives processing tasks from one or more client applications, divides the processing tasks into units of work, and assigns the units of work to one or more of the nodes. Dividing a processing task into units of work may involve dividing audio input data into segments defined by natural speech boundaries. A mathematical representation may be created for each segment prior to its distribution on the grid to minimize network traffic. A node in the system may perform heterogeneous units of work concurrently, such as by isolating the execution of each unit of work in an application domain.