Context-Aware Input Correction for User-Specific Term Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing on-device input recognition systems face challenges in accurately recognizing user-specific terms, such as phonetic variations, misspellings, and personal terms, especially in speech recognition, which can also raise privacy concerns, and existing techniques for improving recognition are not real-time or effective for infrequent or domain-specific terms.
Innovation Solution
A context-aware correction module using a trained machine learning model, like a convolutional neural network, on-device system that identifies and corrects potentially mistranscribed terms by substituting them with more accurate alternatives, based on user-specific data and interactions, maintaining privacy and reducing computational load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If user-specific terms and phonetic variations are added to the recognition model to improve accuracy, then recognition accuracy improves, but device complexity and computational resources increase
Solution Approach 1:
The system segments the recognition model into two parts: a base model for common terms and a user-specific adapter for personalized terms. This allows the heavy lifting to be done by a compact base model while user-specific customizations are handled by a smaller adapter layer, reducing overall complexity while maintaining accuracy for both common and personalized terms.
Solution Approach 2:
The system applies local quality by maintaining a lightweight base model for general recognition while adding localized user-specific adaptations only where needed. User-specific phonetic variations and terms are handled by a dedicated adapter component rather than expanding the entire model, thus improving accuracy for specific cases without proportionally increasing overall complexity.
2Measurement precision
If real-time correction is implemented on-device to improve recognition accuracy, then accuracy improves, but power consumption increases
Solution Approach 1:
The system extracts and isolates the correction functionality into a separate, lightweight component that operates independently from the main recognition model. This allows the correction module to be optimized for low power consumption while the base model remains compact, enabling real-time operation without excessive energy usage.
Solution Approach 2:
The system changes the operational parameters of the correction module to optimize for power efficiency. The adapter uses pre-computed transformations and operates on limited data dimensions, reducing computational load and power consumption while maintaining real-time correction capability for user-specific terms.
3Measurement precision
If a comprehensive speech recognition model is used to recognize domain-specific terms, then recognition accuracy improves, but processing time and latency increase
Solution Approach 1:
The system performs preliminary action by pre-computing and storing user-specific phonetic variations and domain-specific terms in the adapter component during offline training. This allows the runtime system to quickly match incoming speech against pre-prepared patterns without performing complex computations, reducing latency while maintaining high accuracy for specialized terms.
Solution Approach 2:
The system creates simplified copies of user-specific terms and phonetic patterns in the adapter component. These copies are optimized representations that capture essential characteristics without the full complexity of the original data, enabling fast matching and recognition with reduced processing time.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Apparatus and methods related to input recognition are provided. A method includes receiving an input from a user during an interaction with a computing device. The method further includes receiving, from an input recognition model, a transcription of the input. The method also includes identifying a candidate term for substitution in the transcription. The candidate term is likely to have been mistranscribed. The method additionally includes accessing, based on the candidate term, a plurality of pairs of mistranscribed terms and non common terms. The non-common terms are likely to be mistranscribed, and the mistranscribed terms are incorrect versions of the non-common terms, the mistranscribed terms having been generated by a machine learning model. The method further includes substituting, based on the plurality of pairs of mistranscribed terms and non-common terms, the candidate term with a non-common term in the transcription.