Speaker Category Identification for Speech Recognition Accuracy

The patent segments the speech recognition system into multiple specialized engines, each tuned to specific speaker types (e.g., male/female, different accents, age groups). Instead of using a single general-purpose engine, the system divides the recognition task across multiple specialized engines, allowing each to excel at recognizing particular speaker categories while maintaining overall system accuracy.

The patent introduces a speaker type identification module as an intermediary that analyzes incoming speech to determine the speaker's category. This mediator then selects the appropriate specialized recognition engine for processing, avoiding the need to run all engines simultaneously and reducing overall system complexity while maintaining high accuracy for the identified speaker type.

The patent performs preliminary classification of speakers into categories during the enrollment process. Instead of requiring extensive adaptation for each individual speaker, the system pre-categorizes speakers based on acoustic features (gender, accent, age group) and assigns them to appropriate pre-trained recognition engines. This preliminary action significantly reduces the time needed for individual speaker adaptation while maintaining high recognition accuracy.

The patent changes the approach from individual speaker adaptation to category-based parameter selection. By identifying key acoustic parameters that define speaker categories and using these to select appropriate recognition models, the system achieves high accuracy without requiring time-consuming individual enrollment for each user. The parameter change shifts focus from fine-grained individual characteristics to broader categorical features.

The speaker type identification module acts as an energy-efficient intermediary that quickly categorizes incoming speech before routing it to the appropriate specialized engine. This mediation prevents the system from activating multiple heavy computational engines simultaneously, reducing energy consumption while maintaining high recognition accuracy by directing speech to the most suitable engine based on speaker characteristics.

Instead of deploying and running all speech recognition engines for every input, the patent uses partial action by selecting and activating only the specific engine(s) needed for the identified speaker type. This selective activation significantly reduces computational energy usage while maintaining high accuracy, as the system performs just enough processing (using only the necessary engine) rather than excessive processing with all engines.