Speaker Verification Module Parameter Adaptation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Speaker identification systems face reliability issues due to insufficient training data that fails to encompass intra-speaker variability and highlight inter-speaker variability, leading to high false alarm and missed target rates in deployed voice-based interactive systems.

Innovation Solution

The system updates its parameters using actual field audio data, specifically employing i-vectors generated from field data to adapt the i-vector based speaker verification module, allowing for improved recognition accuracy over time with minimal overhead in disk occupancy and computational complexity, while maintaining system security.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If speaker identification systems use parameters generated during pre-deployment enrollment phase, then the system can be initially configured, but the reliability of speaker identification deteriorates due to insufficient training data that fails to encompass intra-speaker variability and highlight inter-speaker variability

Engineering Contradiction:
Improvespeaker identification reliabilityVSAvoidtraining data quantity
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The system performs preliminary classifier adaptation during the enrollment phase using available training data to establish initial speaker parameters. This preliminary configuration enables the system to function immediately while continuing to accumulate field data for future refinement, thus resolving the contradiction between having sufficient training data and maintaining initial system configuration.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system transitions from static parameters established during enrollment to dynamic parameter adaptation using ongoing field data. The classifier is continuously refined using actual audio data collected during deployment, allowing the system to adapt to real-world intra-speaker variability and improve reliability over time without requiring complete retraining.

Inventive Principle:
Principle #15Dynamics

2Measurement precision

If actual field audio data is used to update classifier parameters, then speaker recognition accuracy improves, but system complexity and computational overhead increase

Engineering Contradiction:
Improvespeaker recognition accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system implements a feedback mechanism where classification results from field audio data are fed back to update classifier parameters. The adaptation module uses verification outcomes and accumulated field data to iteratively refine speaker parameters, improving accuracy while maintaining a manageable complexity through automated feedback loops rather than manual system redesign.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The classifier performs self-adaptation using its own accumulated field data and verification results. The system automatically updates its parameters without requiring external intervention or complex external training infrastructure, reducing overall system complexity while continuously improving speaker recognition accuracy through self-learning.

Inventive Principle:
Principle #25Self-service

3Reliability

If classifier parameters are adapted using field data, then false alarm and missed target rates reduce, but data storage requirements increase

Engineering Contradiction:
Improvefalse alarm and missed target ratesVSAvoiddata storage requirements
Core Design Contradiction:
ReliabilityVSVolume of stationary object

Solution Approach 1:

The system extracts only the essential adaptation information from field audio data - specifically, updated speaker parameters and classification statistics - rather than storing all raw field audio. This extraction approach reduces storage requirements while maintaining the ability to improve false alarm and missed target rates through parameter adaptation.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system transforms raw field audio data into condensed parameter representations that capture the essential variability information needed for adaptation. By changing from storing extensive audio data to storing refined parameter sets, the system reduces storage requirements while maintaining improved reliability through accurate speaker parameter updates.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP2770502B1Method and apparatus for automated speaker classification parameters adaptation in a deployed speaker verification system
Publication Date: 2019.11.27 NUANCE COMMUNICATIONS INC
  • EP2770502B1 patent drawingFigure 1
  • EP2770502B1 patent drawingFigure 2
  • EP2770502B1 patent drawingFigure 3

AI summary

Typical speaker verification systems usually employ speakers' audio data collected during an enrollment phase when users enroll with the system and provide respective voice samples. Due to technical, business, or other constraints, the enrollment data may not be large enough or rich enough to encompass different inter-speaker and intra-speaker variations. According to at least one embodiment, a method and apparatus employing classifier adaptation based on field data in a deployed voice-based interactive system comprise: collecting representations of voice characteristics, in association with corresponding speakers, the representations being generated by the deployed voice-based interactive system; updating parameters of the classifier, used in speaker recognition, based on the representations collected; and employing the classifier, with the corresponding parameters updated, in performing speaker recognition.