Hybrid Grammar Statistical Language Model for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic speech recognition systems face challenges in balancing accuracy and performance latency, as grammar-only language models are fast but limited in recognizing deviations, while statistical language models are accurate but slow and less accurate in noisy environments.
Innovation Solution
A combined language model that links grammar-only and statistical language models via backoff arcs, allowing the system to switch between models based on the input, ensuring fast recognition of predefined sequences while accurately handling deviations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If a grammar-only language model is used, then speech recognition speed is improved, but recognition accuracy for deviations from predefined sequences deteriorates
Solution Approach 1:
The patent combines a grammar-only language model and a statistical language model into a single hybrid system. The grammar model provides fast recognition for predefined sequences while the statistical model handles deviations and unknown phrases, resolving the contradiction between speed and accuracy by merging the strengths of both approaches.
Solution Approach 2:
The system dynamically switches between grammar-only mode and statistical model mode based on the input characteristics. For predefined sequences, it uses the fast grammar model; for deviations or unknown phrases, it transitions to the more accurate statistical model, making the system adaptive to different recognition scenarios.
2Reliability
If a statistical language model is used, then recognition accuracy for various word combinations is improved, but speech recognition speed deteriorates
Solution Approach 1:
The patent segments the recognition process into two paths: a fast grammar-based path for predefined sequences and a more comprehensive statistical model path for deviations. This segmentation allows the system to use the appropriate model for each case, maintaining speed for common phrases while achieving accuracy for varied inputs.
Solution Approach 2:
The system applies different model qualities to different recognition scenarios. For predefined sequences, it uses the lightweight grammar model with high speed characteristics. For unknown or deviating phrases, it employs the full statistical model with high accuracy characteristics, optimizing performance locally for each case.
3Loss of time
If a grammar-only language model is used, then performance latency is reduced, but the ability to recognize non-predefined sequences deteriorates
Solution Approach 1:
The patent merges the fast grammar model with the versatile statistical model into a hybrid system. This combination maintains low latency for predefined sequences through the grammar model while gaining the ability to recognize non-predefined sequences through the statistical model component.
4Adaptability or versatility
If the search space is expanded to recognize any word combination, then adaptability is improved, but processing time increases
Solution Approach 1:
The patent segments the search space into a constrained grammar-based search space for fast processing and an expanded statistical model search space for comprehensive coverage. By segmenting the search strategy, the system achieves both adaptability for any word combination and efficient processing time.
Data Source
AI summary
Features are disclosed for performing speech recognition on utterances using a grammar and a statistical language model, such as an n-gram model. States of the grammar may correspond to states of the statistical language model. Speech recognition may be initiated using the grammar. At a given state of the grammar, speech recognition may continue at a corresponding state of the statistical language model. Speech recognition may continue using the grammar in parallel with the statistical language model, or it may continue using the statistical language model exclusively. Scores associated with the correspondences between states (e.g., backoff arcs) may be determined according to a heuristically or based on test data.