Speech Indexing Accuracy via Dual-Engine Comparison

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Efficient Automatic Voice Recognition (AVR) engines used in speech indexing systems often produce false entries due to high error rates, leading to inaccuracies in searchable databases, requiring improvements in accuracy to correctly identify and timestamp key phrases in audio recordings.

Innovation Solution

A dual-indexing approach is employed, where a phonetic automatic voice recognition engine generates a speech index, and a large vocabulary speech recognition (LVSR) engine provides a more accurate secondary index. The results are compared to adjust parameters of the phonetic engine, correcting low-confidence classifications and refining the speech indexing process.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If an efficient phonetic AVR engine is used for speech indexing, then processing speed is improved, but accuracy deteriorates due to higher error rates and false entries

Engineering Contradiction:

Improveprocessing speedVSAvoidaccuracy

Core Design Contradiction:

SpeedVSReliability

Solution Approach 1:

The speech recognition task is segmented into two parts: a fast phonetic engine handles initial processing while a slower LVSR engine handles verification of low-confidence results. This segmentation allows the system to benefit from both speed and accuracy

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

A confidence threshold mechanism acts as an intermediary between the phonetic engine and LVSR engine. Low-confidence results from the phonetic engine are filtered and passed to the LVSR engine for verification, enabling selective use of the slower engine

Inventive Principle:

Principle #24Intermediary (Mediator)

2Reliability

If a more accurate LVSR engine is used for speech indexing, then accuracy is improved, but processing speed deteriorates

Engineering Contradiction:

ImproveaccuracyVSAvoidprocessing speed

Core Design Contradiction:

ReliabilityVSSpeed

Solution Approach 1:

The LVSR engine is not applied to all speech segments but only to portions identified as low-confidence by the phonetic engine. This partial application maintains accuracy where needed while preserving overall processing speed

Inventive Principle:

Principle #16Partial or excessive action

3Reliability

If confidence threshold parameters are adjusted to reduce false entries, then accuracy is improved, but more low-confidence results require LVSR verification

Engineering Contradiction:

ImproveaccuracyVSAvoidtime for LVSR verification

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The system uses feedback from comparing phonetic and LVSR results to dynamically adjust the confidence threshold parameter. This feedback loop optimizes the balance between reducing false entries and minimizing LVSR verification time

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS8326631B1Systems and methods for speech indexing

Publication Date: 2012.12.04 VERINT SYST INC

AI summary

A speech index for a recording or other representation of an audio signal containing speech is generated using a phonetic automatic voice recognition engine. A second speech index is also generated using a more accurate, but slower, automatic voice recognition engine such as a large vocabulary speech recognition (LVSR) engine. These two speech indexes are compared. The results of the comparison are then used to adjust certain parameters used by the phonetic engine while generating a speech index. The results may also be used to correct all or parts of the speech index generated by the phonetic automatic speech recognition engine.