Speech Recognition Error Correction via User Context
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional computer speech recognition (CSR) systems face challenges in accurately recognizing human speech due to variations in speech patterns, such as accent, speed, and health conditions, which are not fully captured by static word models.
Innovation Solution
A computerized method for speech recognition correction is introduced, which involves receiving a voice recognition input from an individual user and using a trained error correction model to add new alternative results to the results list. This model is trained with contextual information including historical user correction logs, personal class definitions, and application context, and a re-ranker is used to re-rank the results list.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If static word models are used for speech recognition, then the system structure is simple, but speech recognition accuracy deteriorates due to inability to capture speech variations
Solution Approach 1:
The patent transforms the static word model into a dynamic system by introducing an error correction model that adapts to individual user speech patterns. The system dynamically adjusts recognition parameters based on user-specific error patterns learned from correction logs, enabling the model to capture speech variations without requiring complete retraining for each user.
Solution Approach 2:
The patent changes the parameters of the recognition system by introducing user-specific error correction parameters. Instead of modifying the entire static word model, the system adjusts specific parameters related to error patterns, accents, and speech variations through the error correction model, improving accuracy while maintaining the simplicity of the base system.
2Measurement precision
If probabilistic techniques are used to improve speech recognition accuracy, then recognition accuracy improves, but errors still occur and the system becomes more complex
Solution Approach 1:
The patent introduces an error correction model as an intermediary component between the voice recognition system and the final output. This intermediary layer processes the initial recognition results and applies user-specific corrections based on learned error patterns, reducing errors without requiring complete redesign of the base recognition system.
Solution Approach 2:
The patent segments the speech recognition process into two distinct stages: initial voice recognition and subsequent error correction. This segmentation allows each component to be optimized independently - the base system handles general recognition while the error correction model focuses specifically on improving accuracy through user-specific adjustments.
3Measurement precision
If training is performed with a large variety of pre-selected words to improve accuracy, then recognition accuracy improves, but the training process becomes time consuming
Solution Approach 1:
The patent performs preliminary error analysis by collecting and analyzing user correction logs before actual speech recognition. The error correction model is pre-trained on user-specific error patterns, allowing the system to quickly adapt to individual users without requiring extensive real-time training during actual usage.
Solution Approach 2:
The system enables self-service learning by automatically collecting user corrections and using them to improve future recognition. The error correction model learns from user feedback without requiring manual retraining, continuously improving accuracy while minimizing user time investment.
Data Source
AI summary
Systems and methods for speech recognition correction include receiving a voice recognition input from an individual user and using a trained error correction model to add a new alternative result to a results list based on the received voice input processed by a voice recognition system. The error correction model is trained using contextual information corresponding to the individual user. The contextual information comprises a plurality of historical user correction logs, a plurality of personal class definitions, and an application context. A re-ranker re-ranks the results list with the new alternative result and a top result from the re-ranked results list is output.