Voice Directory Search Using Hierarchical Language Model Selection

The patent divides the single large speech recognition language model into multiple smaller, specialized language models organized in a hierarchical tree structure. Each node in the tree represents a specific business type or location category and has its own dedicated language model. This segmentation allows the system to select only the relevant language models needed for a particular query, improving recognition accuracy while reducing the computational burden of managing and processing a single enormous model.

The patent implements a hierarchical tree structure where language models are nested at different levels of the hierarchy. General language models at higher levels encompass more specific language models at lower levels. This nesting allows the system to store language models in an organized hierarchy where broader category models contain or reference more specific subset models, reducing redundant storage while maintaining comprehensive coverage for accurate business identification.

The system performs preliminary action by automatically determining the business type category before conducting the speech recognition process. The hierarchical tree structure pre-organizes language models by business type and location, allowing the system to select the appropriate specialized language model in advance based on the query context. This preliminary selection of the relevant language model subset improves recognition accuracy without requiring additional user input, as the system autonomously navigates the hierarchy to find the most appropriate model.