Speech Recognition Correction via Custom Language Model Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computer-implemented speech recognition systems fail to accurately recognize and learn from their mistakes, leading to repeated misinterpretation of speech commands and inefficient user interactions.
Innovation Solution
The implementation of unsupervised learning techniques to detect repeated speech commands and create a customized language model based on clusters of similar predictions, supplemented with text-based search queries, to improve speech recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a universal language model is used for speech recognition, then the system can handle diverse speech inputs, but the accuracy for repeated speech commands deteriorates
Solution Approach 1:
The patent segments the language model processing into two distinct components: a universal language model for general speech recognition and a custom language model for repeated speech commands. This segmentation allows each model to specialize in its respective domain, with the universal model maintaining versatility and the custom model improving accuracy for repeated commands through clustered predictions from the universal model's n-best list.
Solution Approach 2:
The system dynamically switches between the universal language model and the custom language model based on whether a speech command is repeated. The custom language model is activated specifically for repeated commands within a session, allowing the system to adapt its processing approach based on the nature of the input while maintaining the universal model for diverse, non-repeated inputs.
2Measurement precision
If the speech recognition processor retries recognition for repeated commands, then accuracy may improve, but processing time and computational resources increase
Solution Approach 1:
The system performs preliminary clustering of n-best predictions from the universal language model to identify repeated speech commands before final recognition. By detecting patterns in repeated commands upfront and preparing the custom language model in advance, the system avoids multiple full recognition cycles, reducing processing time while maintaining improved accuracy through the custom model's focused predictions.
3Measurement precision
If the speech recognition system uses a custom language model for repeated commands, then recognition accuracy improves, but device complexity increases
Solution Approach 1:
The patent introduces an intermediary component that bridges the universal language model and the custom language model. This intermediary processes the n-best predictions from the universal model, clusters them to identify repeated commands, and generates the custom language model on-demand. This intermediary layer manages the complexity by automating the creation and switching between models, preventing the system from becoming overly complex while maintaining the benefits of customized recognition.
Data Source
AI summary
Disclosed in various examples are methods, systems, and machine-readable mediums for providing improved computer implemented speech recognition by detecting and correcting speech recognition errors during a speech session. The system recognizes repeated speech commands from a user in a speech session that are similar or identical to each other. To correct these repeated errors, the system creates a customized language model that is then utilized by the language modeler to produce a refined prediction of the meaning of the repeated speech commands. The custom language model may comprise clusters of similar past predictions of speech commands from the speech session of the user.