Personalized Speech Recognition Model Update
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition technologies often fail to accurately recognize user speech, leading to user dissatisfaction and reduced performance due to misrecognition, which is not effectively improved by reutterance, as the devices lack personalized adaptation to individual speech patterns.
Innovation Solution
An artificial intelligence apparatus and method that utilizes personalized language models, updating them based on user identification information and speech recognition results to improve speech recognition accuracy by adjusting word weights and correcting misrecognized words using a combination of global and personalized language models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a global language model is used for speech recognition, then the system can recognize speech from any user, but the recognition accuracy drops for individual users due to lack of personalization
Solution Approach 1:
The language model is divided into two segments: a global language model that handles general speech recognition for all users, and personalized language models that are segmented by user identification information to handle individual speech patterns. This segmentation allows the system to maintain versatility while improving accuracy for each user through personalized modeling.
Solution Approach 2:
The system merges the global language model with personalized language models by combining their probability scores. When recognizing user speech, the system integrates results from both the global model (providing broad coverage) and the personalized model (providing user-specific accuracy), thereby resolving the contradiction between versatility and precision.
2Reliability
If speech recognition fails and the user is asked to reutter, then the system maintains its current model, but the performance does not improve because the device lacks personalized adaptation
Solution Approach 1:
The system implements feedback by using speech recognition results to update personalized language models. When speech is recognized (even with minor errors), the system feeds this information back to refine the personalized model for that user, enabling continuous performance improvement without requiring explicit user correction input.
Solution Approach 2:
The system performs self-service by automatically updating its personalized language models using its own speech recognition data. The device learns from its recognition attempts and automatically adapts to each user's speech patterns without external intervention, thereby improving productivity and performance over time.
3Measurement precision
If the system uses a personalized language model for each user, then speech recognition accuracy improves, but the system complexity increases due to multiple models
Solution Approach 1:
The personalized language model system is designed to be universal by using a common model structure and update mechanism for all users. Each user has a personalized model, but they all follow the same architectural blueprint and update process, reducing the effective complexity while maintaining personalized accuracy.
Solution Approach 2:
The system manages complexity by changing parameters rather than restructuring the entire model. Personalization is achieved by adjusting specific parameters (word weights, probabilities) in the language model based on user data, rather than creating entirely different model architectures for each user, thereby maintaining simplicity while improving accuracy.
4Measurement precision
If misrecognized words are updated in the personalized language model, then recognition accuracy for those words improves, but the model requires more computational resources for updating
Solution Approach 1:
The system applies partial action by updating only the specific words that were misrecognized rather than retraining the entire language model. This selective updating approach improves accuracy for problematic words while minimizing computational energy expenditure, avoiding the excessive action of complete model retraining.
Data Source
AI summary
An artificial intelligence apparatus for recognizing speech of a user includes a microphone, and a processor configured to receive, via the microphone, a sound signal corresponding to the speech of the user, acquire personalize identification information corresponding to the speech, recognize the speech from the sound signal using a global language model, calculate a reliability for the recognition, and if the calculated reliability exceeds a predetermined first reference value, update a personalized language model corresponding to the personalize identification information using the recognition result.