Selectively Biased LDA for Speech Recognition Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Automatic speech recognition systems face challenges in maintaining accuracy and reliability due to factors like accent, articulation, speech rate, and background noise, especially in varying channel conditions such as VoIP, which affect their performance and processing speed.

Innovation Solution

The implementation of selectively biased linear discriminant analysis (LDA) enhances the discrimination between HMM tied-states in the acoustic feature space by biasing between-class and within-class covariance matrices based on observed recognition errors, using a two-stage process involving orthogonal transformation and principal component analysis to improve feature extraction and classification.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional linear discriminant analysis is used for feature extraction in speech recognition, then the system can process speech data, but recognition accuracy deteriorates under varying channel conditions such as noise and different speech characteristics

Engineering Contradiction:
Improverecognition accuracyVSAvoidchannel conditions
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent applies local quality by selectively biasing the within-class scatter matrix for specific tied-states based on their individual confusion patterns. Instead of applying a uniform transformation to all features, the method identifies which tied-states contribute most to recognition errors and applies enhanced discrimination only to those specific regions in the feature space. This targeted approach improves robustness to channel variations without unnecessarily complicating the overall system.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent modifies the standard LDA parameters by introducing a bias term to the within-class scatter matrix calculation. Specifically, it adds a scaled version of the identity matrix to the within-class scatter matrix for selected tied-states, changing the parameter structure to emphasize separation in directions that are most problematic for recognition accuracy under varying conditions.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If selective biasing of covariance matrices is applied to improve discrimination between HMM tied-states, then recognition accuracy improves, but computational complexity increases

Engineering Contradiction:
Improverecognition accuracyVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the speech recognition problem by identifying and treating individual tied-states separately. Rather than applying complex transformations to the entire feature space, it divides the problem into manageable portions by focusing on specific tied-states that contribute most to recognition errors. This segmentation allows the system to apply computational resources selectively, improving accuracy while controlling overall complexity.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies partial action by selectively biasing only certain tied-states rather than all tied-states uniformly. The method identifies which tied-states have high confusion rates and applies the biased LDA transformation only to those specific states, leaving other tied-states to be processed with standard LDA or simpler transformations. This partial application reduces computational burden while still achieving improved recognition accuracy.

Inventive Principle:
Principle #16Partial or excessive action

3Reliability

If feature extraction is enhanced to reduce recognition errors, then processing time increases, but real-time performance requirements must be maintained

Engineering Contradiction:
Improverecognition accuracyVSAvoidprocessing speed
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent performs preliminary action by pre-computing the biased scatter matrices and transformation parameters during the training phase. By calculating and storing the biased within-class scatter matrices for each tied-state in advance, the system avoids performing complex matrix operations during real-time speech recognition. This preliminary computation shifts the computational burden to the offline training stage, allowing faster processing during actual recognition while maintaining improved accuracy.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP2888669B1Method and system for selectively biased linear discriminant analysis in automatic speech recognition systems
Publication Date: 2021.03.31 INTERACTIVE INTELLIGENCE INC
  • EP2888669B1 patent drawingFigure 1
  • EP2888669B1 patent drawingFigure 2
  • EP2888669B1 patent drawingFigure 3

AI summary

A system and method are presented for selectively biased linear discriminant analysis in automatic speech recognition systems. Linear Discriminant Analysis (LDA) may be used to improve the discrimination between the hidden Markov model (HMM) tied-states in the acoustic feature space. The between-class and within-class covariance matrices may be biased based on the observed recognition errors of the tied-states, such as shared HMM states of the context dependent tri-phone acoustic model. The recognition errors may be obtained from a trained maximum-likelihood acoustic model utilizing the tied-states which may then be used as classes in the analysis.