Semantic Confidence Measure for ASR Error Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing confidence measure calculation methods in speech recognition systems may incorrectly assign high confidence scores to recognition errors due to large semantic relatedness between target words and adjacent context, leading to incorrect judgment of recognition accuracy.
Innovation Solution
A method and apparatus that generate alternative candidate word strings, train classifiers to discriminate between target and candidate word strings, and calculate confidence measures based on the classification results, ensuring accurate detection of recognition errors even when semantic relatedness is high.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If semantic relatedness between target word and adjacent context is used to calculate confidence measure, then ease of calculation is improved, but measurement precision deteriorates because recognition errors with high semantic relatedness are incorrectly assigned high confidence scores
Solution Approach 1:
The patent segments the confidence measure calculation into multiple independent components: phoneme-level matching score, word-level semantic relatedness score, and context-level coherence score. Each component is calculated separately and then integrated, allowing the system to capture both local pronunciation accuracy and global semantic consistency without compromising computational efficiency
Solution Approach 2:
The patent introduces an intermediary confidence measure calculation unit that acts as a mediator between the acoustic model and language model. This unit integrates multiple features including phoneme matching scores, semantic relatedness, and context coherence, producing a comprehensive confidence measure that resolves the contradiction between calculation simplicity and measurement accuracy
2Speed
If simple semantic relatedness calculation is used, then calculation speed is improved, but reliability deteriorates due to incorrect error detection
Solution Approach 1:
The confidence measure calculation is segmented into parallel independent computations: phoneme matching at the acoustic level, semantic relatedness at the word level, and context coherence at the sentence level. This segmentation enables simultaneous calculation of multiple features without sequential bottlenecks, maintaining high speed while improving reliability through multi-dimensional analysis
Solution Approach 2:
The patent extends the confidence measure from a single scalar value to a multi-dimensional vector encompassing phoneme-level accuracy, word-level semantics, and context-level coherence. This dimensional expansion allows the system to detect recognition errors that might be missed by simple semantic relatedness alone, thereby improving reliability without sacrificing calculation speed
Data Source
AI summary
An apparatus that calculates a confidence measure of a target word string specified in a recognition result includes: an alternative candidate generator which generates an alternative candidate word string in the position of the target word string; a classifier training unit which trains a classifier which is configured to discriminate between the target word string and the alternative candidate word string; a feature extractor which extracts a feature value representing an adjacent context in the position of the target word string; and a confidence measure calculator which determining whether the true word string in the position of the target word string is the target word string or the alternative candidate word string by using the classifier and the feature value and calculates a confidence measure of the target word string on the basis of the determination result.