Speech Recognition Grammar Expansion via Statistical Language Model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face limitations in recognizing and interpreting speech inputs that fall outside the predefined grammar, leading to reduced accuracy and inability to differentiate between gibberish and linguistically well-formed speech, especially when users provide inputs not covered by the developer-provided grammar.
Innovation Solution
The implementation enhances speech recognition by using a statistical language model that includes annotations providing syntax and semantics, which processes user input to generate a wider range of candidates while preserving the underlying context and semantics, allowing for improved recognition and interpretation of speech inputs through weighting and expansion of the base speech recognition grammar.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a base speech recognition grammar is used with predefined syntax and semantics, then the system structure is simple and easy to construct, but the system cannot recognize speech inputs that fall outside the predefined grammar
Solution Approach 1:
The patent merges a base speech recognition grammar with a statistical language model to create an enhanced recognition system. The base grammar provides structured syntax and semantics while the statistical model expands coverage to handle unseen speech patterns, combining the advantages of both predefined and adaptive approaches
Solution Approach 2:
The enhanced speech recognition system achieves multi-functionality by integrating multiple components: the base grammar for structured recognition, the statistical language model for expanded coverage, and candidate generation mechanisms. This universal system can handle both predefined and unexpected speech inputs
2Adaptability or versatility
If the speech recognition grammar is expanded to cover more speech inputs, then the recognition coverage increases, but the system complexity increases
Solution Approach 1:
The patent segments the speech recognition system into distinct functional components: the base grammar module, the statistical language model module, and the candidate generation module. This segmentation allows each component to be optimized independently while working together to achieve comprehensive coverage without overwhelming system complexity
Solution Approach 2:
The statistical language model acts as an intermediary between the base grammar and the speech input. It bridges the gap by generating candidate transcripts that connect predefined grammar structures with unseen speech patterns, expanding coverage without directly increasing grammar complexity
3Measurement precision
If a statistical language model is used to generate more candidates, then the recognition accuracy improves, but the processing time increases
Solution Approach 1:
The system applies partial action by generating a limited set of high-probability candidate transcripts from the statistical language model rather than exhaustively evaluating all possible transcripts. This selective approach maintains high recognition accuracy while reducing processing time through focused candidate generation
4Reliability
If the base grammar is weighted using statistical speech recognition information, then the recognition robustness improves, but the computational resources required increase
Solution Approach 1:
The patent changes the parameter of grammar candidates by assigning statistical weights to them based on speech recognition information. This parameter modification enhances recognition robustness by prioritizing more likely candidates while managing computational resources through weighted evaluation rather than exhaustive processing
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for recognizing speech input. In one aspect, a method includes receiving a user input and a grammar including annotations, the user input comprising audio data and the annotations providing syntax and semantics to the grammar, retrieving third-party statistical speech recognition information, the statistical speech recognition information being transmitted over a network, generating a statistical language model (SLM) based on the grammar and the statistical speech recognition information, the SLM preserving semantics of the grammar, processing the user input using the SLM to generate one or more results, comparing the one or more results to candidates provided in the grammar, identifying a particular candidate of the grammar based on the comparing, and providing the particular candidate for input to an application executed on a computing device.