Speech Recognition Entity Scoring for Vertical Domain Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In vertical or professional speech recognition scenarios, specific entities often have low recognition accuracy due to differences from ordinary semantic logic, which existing speech recognition technologies fail to address effectively.
Innovation Solution
A method that acquires candidate recognition results from a speech recognition model, scores them based on pronunciation similarities with pre-collected popular entities, and integrates these scores to determine a final recognition result, improving accuracy by aligning with popular entities' pronunciation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a traditional speech recognition model is used to recognize speech, then the recognition process is simple and fast, but the recognition accuracy is low for specific entities in vertical or professional scenarios
Solution Approach 1:
The patent introduces an entity scoring module as an intermediary component between the speech recognition model and the final result selection. This module calculates entity scores for candidate recognition results based on entity knowledge, acting as a mediator that bridges the gap between traditional speech recognition and improved accuracy for specific entities. The entity scoring module processes candidate results and provides additional scoring information without completely replacing the original recognition system.
Solution Approach 2:
The patent merges the traditional speech recognition scoring mechanism with a new entity scoring mechanism. The final recognition result is determined by combining the recognition score from the speech recognition model and the entity score from the entity scoring module. This merging of multiple scoring dimensions allows the system to leverage both the speed and general accuracy of traditional recognition and the entity-specific accuracy of the new module.
2Measurement precision
If entity scoring is introduced to improve recognition accuracy for specific entities, then recognition accuracy improves, but the computational complexity and processing time increase
Solution Approach 1:
The patent applies partial action by not requiring entity scoring for all candidate recognition results in all scenarios. The entity scoring module can be selectively applied based on confidence thresholds, entity types, or scenario-specific requirements. This partial application reduces unnecessary computational overhead while maintaining accuracy improvements where they are most needed.
Solution Approach 2:
The patent implements local quality by applying entity-specific scoring only to relevant candidate recognition results rather than uniformly to all inputs. The entity scoring mechanism focuses computational resources on candidates that are likely to be specific entities based on initial recognition, applying enhanced processing only where it will have the most impact on accuracy.
3Measurement precision
If multiple scoring dimensions are integrated to determine the final recognition result, then the recognition accuracy for specific entities improves, but the system complexity increases
Solution Approach 1:
The patent designs the entity scoring module with universal applicability across different speech recognition scenarios. The same entity scoring mechanism can be applied to various types of entities (locations, persons, organizations, products) and different speech recognition contexts without requiring scenario-specific customization. This multi-functionality reduces overall system complexity by using a unified approach rather than multiple specialized modules.
Data Source
AI summary
The present disclosure discloses a speech recognition method and apparatus, and relates to the field of speech and deep learning technologies. A specific implementation scheme involves: acquiring candidate recognition results with first N recognition scores outputted by a speech recognition model for to-be-recognized speech, N being a positive integer greater than 1; scoring the N candidate recognition results based on pronunciation similarities between candidate recognition results and pre-collected popular entities, to obtain similarity scores of the candidate recognition results; and integrating the recognition scores and the similarity scores of the candidate recognition results to determine a recognition result corresponding to the to-be-recognized speech from the N candidate recognition results. The present disclosure can improve recognition accuracy.