N-best list rescoring for speech recognition latency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems face challenges in efficiently rescoring recognition hypotheses, particularly in handling large lists of place names, where long latencies and loading times can be problematic, and existing methods do not effectively prioritize frequently used versus less frequently used names.
Innovation Solution
The system rescoring technique involves categorizing recognition hypotheses based on their rank in the N-best list and applying biases or penalties to prioritize recently and frequently used names, allowing for efficient processing and display of top results while deferring less important hypotheses, using a combination of Statistical Language Models and Dynamic Semantic Models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If all recognition hypotheses are processed and displayed in the N-best list, then completeness of recognition results is improved, but user response time deteriorates due to long latencies
Solution Approach 1:
The N-best list is segmented into multiple pages, with only the first page (containing top-ranked hypotheses) displayed initially. This segmentation allows the system to present a manageable subset of results quickly, improving user response time while preserving access to complete results through pagination.
Solution Approach 2:
The system performs preliminary processing to generate and rank all recognition hypotheses before displaying them. By pre-computing the ranked N-best list and preparing pagination structures in advance, the system reduces latency when users request results, as the sorting and ranking work is already completed.
2Measurement precision
If a large vocabulary of place names is loaded for recognition, then recognition accuracy is improved, but peak RAM requirements worsen
Solution Approach 1:
The large place name vocabulary is segmented into multiple data structures including a full vocabulary for accurate recognition and a reduced vocabulary for faster processing. This segmentation allows the system to maintain high recognition accuracy using the full vocabulary while reducing peak memory usage by using the reduced vocabulary during active recognition operations.
Solution Approach 2:
The patent extracts frequently used place names into a separate reduced vocabulary that can be quickly loaded into memory. This extraction allows the system to maintain a compact, fast-access vocabulary for common names while keeping the complete vocabulary available on disk or in slower memory, thereby reducing peak RAM requirements while preserving recognition accuracy for both common and rare place names.
3Measurement precision
If N-best list rescoring is performed using sophisticated models, then recognition accuracy is improved, but computation time worsens
Solution Approach 1:
The system applies rescoring using sophisticated models selectively rather than uniformly to all hypotheses. By focusing computational resources on rescoring only the top-ranked hypotheses or those that meet certain criteria, the system achieves improved recognition accuracy for the most likely candidates while limiting the total computation time spent on rescoring.
Data Source
AI summary
A method of speech recognition processing is described based on an N-best list of recognition hypotheses corresponding to a spoken input. Each hypothesis on the N-best list is rescored based on its rank in the rescored N-best list. The rescoring may be based on a Statistical Language Model (SLM) or Dynamic Semantic Model (DSM). One or more rescoring categories may be associated with each recognition hypotheses to affect or bias the rescoring.