Context-Specific Grammar Recognition via Parallel Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition systems face suboptimal performance in large grammar search spaces, characterized by long recognition latency, poor accuracy, and insufficient grammar coverage, particularly in tasks like voice search that require extensive domain coverage.
Innovation Solution
The proposed architecture employs context-specific grammars for parallel recognition processing, where multiple recognitions are run against smaller, specific grammars, and the results are reconciled using dynamic grammar composition or statistical classification to improve latency, accuracy, and coverage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a large grammar search space is used to cover extensive domain, then grammar coverage is improved, but recognition latency increases and accuracy deteriorates
Solution Approach 1:
The patent divides a large grammar search space into multiple smaller context-specific grammars (e.g., business names grammar, web search queries grammar, voice dialing requests grammar). Each grammar handles a specific context or domain, allowing parallel recognition processing that reduces latency while maintaining comprehensive coverage through the union of all grammars.
Solution Approach 2:
The patent introduces a new dimension of context-specific processing by creating multiple recognition paths, each associated with specific context information (such as task type, geographic location, or user preferences). This dimensional expansion allows the system to process queries in parallel across different contexts rather than sequentially through a single large grammar.
2Adaptability or versatility
If a large grammar search space is used to cover extensive domain, then grammar coverage is improved, but recognition accuracy deteriorates
Solution Approach 1:
By segmenting the large grammar into smaller context-specific grammars, each grammar can be optimized for its specific domain, improving recognition accuracy within that context. The system maintains high accuracy by ensuring each segmented grammar is precise for its intended purpose while collectively covering the entire domain.
Solution Approach 2:
The patent applies local quality by making each grammar context-specific, where each grammar is tailored with appropriate constraints and vocabulary for its specific context (e.g., business names grammar contains business-related terms and structures). This localized optimization ensures high accuracy for context-appropriate queries while maintaining overall system versatility.
3Loss of time
If multiple context-specific grammars are used in parallel, then recognition latency is reduced, but system complexity increases
Solution Approach 1:
The system segments recognition into multiple parallel paths, each handling a specific context. This segmentation enables concurrent processing that reduces latency while the modular structure manages complexity by organizing grammars and recognizers into independent, manageable units that can be selectively activated based on context.
Solution Approach 2:
The patent implements preliminary action by pre-organizing grammars into context-specific categories and pre-determining which grammars should be activated for different query types. This preliminary structuring allows the system to quickly select and execute appropriate recognition paths without complex real-time decision-making, reducing operational complexity while maintaining parallel processing benefits.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Architecture that employs an overall grammar as a set of context-specific grammars for recognition of an input, each responsible for a specific context, such as subtask category, geographic region, etc. The grammars together cover the entire domain. Moreover, multiple recognitions can be run in parallel against the same input, where each recognition uses one or more of the context-specific grammars. The multiple intermediate recognition results from the different recognizer-grammars are reconciled by running re-recognition using a dynamically composed grammar based on the multiple recognition results and potentially other domain knowledge, or selecting the winner using a statistical classifier operating on classification features extracted from the multiple recognition results and other domain knowledge.