Active Speaker Recognition via Pre-trained Voiceprint Database

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing systems for detecting and recognizing active speakers in teleconferences lack accuracy, replicability, and robustness, especially when video is not available, due to high computational complexity.

Innovation Solution

A method and system that trains a Convolutional Neural Network (CNN) to calculate reference vectors from voice samples before a teleconference, creating a database of candidate participants, and compares these vectors with real-time voice stream vectors during the conference to identify active speakers efficiently.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If Gaussian Mixture Model or Neural Networks are used for active speaker detection, then detection capability is provided, but computational complexity increases and processing time becomes high

Engineering Contradiction:
Improvedetection capabilityVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system performs preliminary action by extracting and storing voiceprint features of known participants before the conference starts. These pre-extracted features are stored in a database and used during the conference to quickly identify speakers, avoiding the need for complex real-time neural network processing of all audio data.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The invention extracts only the essential voiceprint features from the audio signals and stores them in a database. During the conference, only these extracted features are compared against the database rather than processing the entire audio stream through complex models, thereby reducing computational complexity while maintaining detection capability.

Inventive Principle:
Principle #2Taking out (Extraction)

2Measurement precision

If complex neural network models are used for speaker recognition, then recognition accuracy may improve, but processing time increases significantly

Engineering Contradiction:
Improverecognition accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system creates a database of pre-computed voiceprint features that copy the essential characteristics of speakers' voices. During the conference, the system compares incoming audio features against these stored copies using simple similarity matching algorithms, achieving fast processing times while maintaining accurate speaker recognition without requiring complex real-time neural network inference.

Inventive Principle:
Principle #26Copying

3Reliability

If existing speaker detection systems are used, then some detection function is provided, but accuracy and replicability are insufficient

Engineering Contradiction:
Improvedetection functionVSAvoidaccuracy and replicability
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The system changes the parameter representation by transforming audio signals into voiceprint feature vectors that capture essential speaker characteristics. These feature vectors are stored in a database and used for comparison during conferences. This parameter transformation enables more accurate and replicable speaker identification by focusing on distinctive acoustic features rather than raw audio signals.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4040435B1Method and system for teleconference actual participant recognition
Publication Date: 2024.01.17 ALE INT
  • EP4040435B1 patent drawingFigure 1a~1b
  • EP4040435B1 patent drawingFigure 2
  • EP4040435B1 patent drawingFigure 3

AI summary

A method and system for detecting and recognizing an actual participant (607) as an active speaker in a teleconference (606) by training an encoder before the teleconference takes place, and creating a database (602) of reference vectors representing the voice of candidate participants, then, while the teleconference takes place, comparing reference vectors with vectors representing the voice stream of actual participants. The encoder may for example be a Convolutional Neural Network (CNN).