Speech Recognition Word Confidence Score Threshold Tuning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems suffer from low accuracy in word recognition, with existing methods failing to effectively reduce the word recognition error rate.
Innovation Solution
The method involves using word confidence scores (WCS) to modify hypothetical words (mHYP) by determining insertion and substitution threshold values, and employing a selected word list to minimize the weighted total error rate through a tuning phase and testing phase algorithm, which includes converting HYP to mHYP based on WCS and error rate analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional speech recognition decoding is used, then the system operates with standard complexity, but the word recognition accuracy remains low with high error rates
Solution Approach 1:
The patent applies preliminary action by performing a tuning phase before actual speech recognition operations. During this tuning phase, the system pre-calculates optimal insertion and substitution threshold values based on training data, creating lookup tables that store pre-determined threshold settings. This preliminary preparation enables the main recognition system to operate efficiently without real-time complex calculations, thereby improving word recognition accuracy while maintaining manageable system complexity during actual operation.
Solution Approach 2:
The patent replaces complex real-time mechanical decision-making in speech recognition with a data-driven approach. Instead of relying solely on traditional acoustic and language models to make insertion/substitution decisions, the system substitutes these with pre-computed threshold values derived from statistical analysis of error patterns. This substitution transforms the recognition process into a comparison-based system that achieves higher accuracy without proportionally increasing computational complexity.
2Measurement precision
If word confidence scores are used to modify all hypothetical words, then recognition accuracy improves, but processing time and computational resources increase
Solution Approach 1:
The patent applies local quality by differentiating the processing applied to different words based on their characteristics. Instead of uniformly processing all hypothetical words, the system identifies specific words that benefit most from modification (those with ambiguous confidence scores near threshold values) and applies modification selectively. Words with clear high confidence scores are left unchanged, thereby improving recognition accuracy for problematic cases while minimizing unnecessary processing time for confident recognitions.
Solution Approach 2:
The patent implements partial action by applying word modification only when necessary rather than to all words. The system uses confidence score thresholds to determine which words require modification, applying the modification process selectively to words that fall into ambiguous confidence ranges. This partial application of the modification process reduces overall computational burden and processing time while still achieving significant accuracy improvements for the critical subset of problematic words.
3Reliability
If insertion and substitution thresholds are determined through extensive tuning, then error rate decreases, but the tuning phase complexity and data requirements increase
Solution Approach 1:
The patent applies segmentation by breaking down the complex threshold determination process into distinct, manageable components. The tuning phase is segmented into separate tasks: collecting error data from initial recognition attempts, analyzing substitution error patterns, analyzing insertion error patterns, and separately determining optimal thresholds for each type of error. This segmentation of the tuning process makes it more systematic and manageable, reducing the perceived complexity while achieving comprehensive error rate reduction through targeted threshold optimization.
Data Source
AI summary
A method and system for improving the accuracy of a speech recognition system using word confidence score (WCS) processing is introduced. Parameters in a decoder are selected to minimize a weighted total error rate, such that deletion errors are weighted more heavily than substitution and insertion errors. The occurrence distribution in WCS is different depending on whether the word was correctly identified and based on the type of error. This is used to determine thresholds in WCS for insertion and substitution errors. By processing the hypothetical word (HYP) (output of the decoder), a mHYP (modified HYP) is determined. In some circumstances, depending on the WCS's value in relation to insertion and substitution threshold values, mHYP is set equal to: null, a substituted HYP, or HYP.