ASR Confidence Checking for Speech Query Accuracy

Solution Overview

Problem

Current speech processing systems face limitations in answering user queries due to errors in automatic speech recognition (ASR) and natural language understanding (NLU) processing, particularly when relying on structured knowledge bases, which can be incomplete and require human annotation, and automated data acquisition methods are prone to inaccuracies.

Innovation Solution

A system that uses confidence checks on ASR output and compares queries to determine semantic similarity, allowing for the supplementation of knowledge bases with unstructured data from web searches, utilizing machine-learning trained classifiers to identify subject-object-relation triples and improve query answering accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If structured knowledge bases are used for query answering, then query answering can be performed, but the knowledge base becomes incomplete and requires human annotation

Engineering Contradiction:

Core Design Contradiction:

Extent of automationVSReliability

Solution Approach 1:

The system automatically supplements its own knowledge base by extracting information from unstructured web sources through automated techniques, eliminating the need for external human annotation and enabling self-improvement

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent replaces manual human annotation processes with automated machine-learning-based extraction systems that can process unstructured data from web sources to automatically update the knowledge base

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Quantity of substance

If automated data acquisition methods are used, then knowledge base can be expanded, but inaccuracies are introduced

Engineering Contradiction:

Core Design Contradiction:

Quantity of substanceVSMeasurement precision

Solution Approach 1:

The system incorporates confidence checks that provide feedback on the quality of extracted data, allowing the system to verify accuracy and adjust its data acquisition processes accordingly

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent uses confidence thresholds and quality parameters to filter and validate extracted data, changing the acceptance criteria for incoming data based on measured confidence levels

Inventive Principle:

Principle #35Parameter changes

3Measurement precision

If confidence checks are performed on ASR output, then processing accuracy is improved, but processing time increases

Engineering Contradiction:

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system performs confidence checks selectively rather than universally, applying verification only when necessary based on confidence thresholds to balance accuracy improvement with time constraints

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS10332508B1Confidence checking for speech processing and query answering

Publication Date: 2019.06.25

AI summary

An automatic speech recognition (ASR) system uses recurrent neural network (RNN) encoding to create a feature vector corresponding to a word sequence ASR result where the feature vector incorporates data from different hierarchies (i.e., frame level, phone level, etc.) of the ASR processing. The feature vector may be used with a trained classifier to confirm that the ASR result was correct, or to otherwise assign a confidence score to the ASR results.