Speech Recognition Confidence Classifier Score Normalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The accuracy of speech recognition confidence classifier scores becomes obsolete when speech recognition engines or acoustic models are updated, leading to changes in correct-accept versus false-accept profiles that invalidate or render inaccurate the preset operating thresholds in application software.
Innovation Solution
Normalization of speech recognition confidence classifier scores is achieved through histogram-mapping, polynomial-mapping, or tan h-mapping, which maps scores from an old model to a new model, ensuring that preset operating thresholds remain valid and accurate even after updates to the speech recognition engine or acoustic models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If speech recognition engines or acoustic models are updated, then recognition accuracy and performance are improved, but the accuracy of preset operating thresholds becomes obsolete or inaccurate
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the operating threshold based on the confidence score distribution of the updated acoustic model. Instead of using a fixed preset threshold, the system adapts the threshold parameter to match the new model's characteristics, thereby maintaining threshold accuracy despite model updates. This is achieved through analyzing the confidence score histogram and selecting an optimal threshold that preserves the desired balance between correct accepts and false accepts.
2Ease of manufacture
If fixed operating thresholds are used in application software, then software implementation is simplified, but the thresholds become invalid when speech recognition models are updated
Solution Approach 1:
The patent implements dynamics by transforming the static fixed threshold into a dynamic adaptive threshold. The system continuously monitors the confidence score distribution from the speech recognition engine and adjusts the operating threshold accordingly. This dynamic approach allows the threshold to adapt to model updates while maintaining the simplicity of the software implementation, as the adaptation is performed automatically through histogram analysis and optimal threshold selection.
3Productivity
If updates to speech recognition engines are frequent, then recognition performance improves continuously, but recalibration of operating thresholds becomes necessary frequently
Solution Approach 1:
The patent applies self-service by enabling the system to automatically recalibrate the operating threshold without requiring manual intervention. When an acoustic model update occurs, the system autonomously analyzes the new confidence score distribution, determines the optimal threshold, and applies it. This self-calibrating mechanism eliminates the need for developers or operators to perform time-consuming manual recalibration, allowing frequent model updates to occur without proportionally increasing recalibration time.
Data Source
AI summary
The described technology provides normalization of speech recognition confidence classifier (CC) scores that maintains the accuracy of acceptance metrics. A speech recognition CC scores quantitatively represents the correctness of decoded utterances in a defined range (e.g., [0,1]). An operating threshold is associated with a confidence classifier, such that utterance recognitions having scores exceeding the operating threshold are deemed acceptable. However, when a speech recognition engine, an acoustic model, and/or other parameters are updated by the platform, the correct-accept (CA) versus false-accept (FA) profile can change such that the application software's operating threshold is no longer valid or as accurate. Normalizing of speech recognition CC scores to map to the same or better CA and/or FA profiles at the previously-set operating thresholds allows preset operating thresholds to remain valid and accurate, even after a speech recognition engine, acoustic model, and/or other parameters are changed.