Speaker Identification Accuracy via Audio Slicing and Embedding Selection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speaker identification systems face challenges in accurately verifying user identity based on limited or suboptimal audio samples, leading to potential misidentification and unauthorized access.

Innovation Solution

A method and system that generate multiple acoustic embeddings from audio samples by dividing them into slices, applying augmentation techniques, and selecting the most accurate representations to compare and verify speaker identity, using both vector-based and model-based approaches.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If speaker identification systems use limited or suboptimal audio samples, then the system can operate with less data, but the accuracy of user identity verification deteriorates

Engineering Contradiction:
Improvespeaker identification accuracyVSAvoidamount of audio data
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The audio sample is divided into multiple smaller slices, and multiple acoustic embeddings are generated from these slices. This segmentation allows the system to extract multiple representative features from a limited audio sample, improving identification accuracy without requiring more input data.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Multiple acoustic embeddings are generated as copies or representations of the original audio sample's acoustic features. By creating multiple embeddings from the same audio sample, the system increases the amount of useful information extracted without actually increasing the input audio data quantity.

Inventive Principle:
Principle #26Copying

2Reliability

If speaker identification systems use limited audio samples, then data processing time is reduced, but misidentification increases

Engineering Contradiction:
Improvespeaker verification accuracyVSAvoidprocessing time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The audio sample is pre-divided into slices and multiple acoustic embeddings are pre-generated before the actual identification comparison. This preliminary processing allows for more accurate comparisons during the verification process without significantly increasing total processing time.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

Multiple acoustic embeddings are created as efficient representations (copies) of the audio sample's acoustic characteristics. These embeddings can be quickly compared during verification, providing high accuracy with minimal processing time compared to processing the full audio sample directly.

Inventive Principle:
Principle #26Copying

3Reliability

If speaker identification systems use limited audio samples, then storage requirements are reduced, but identification reliability deteriorates

Engineering Contradiction:
Improvespeaker identification accuracyVSAvoidstorage capacity
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

Instead of storing the full audio sample, the system stores multiple acoustic embeddings which are compressed vector representations of the audio's acoustic features. These embeddings capture the essential identification information with much smaller storage requirements while maintaining or improving identification accuracy.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system transforms the audio sample from its original form into acoustic embedding vectors, changing the representation parameters. This transformation reduces the data size significantly while preserving the essential acoustic characteristics needed for accurate speaker identification.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4218009B1Speaker identification accuracy
Publication Date: 2025.06.25 GOOGLE LLC
  • EP4218009B1 patent drawingFigure 1A
  • EP4218009B1 patent drawingFigure 1B
  • EP4218009B1 patent drawingFigure 2A

AI summary

A method (300) of generating an accurate speaker representation for an audio sample (202) includes receiving a first audio sample from a first speaker (10) and a second audio sample from a second speaker. The method includes dividing a respective audio sample into a plurality of audio slices (214). The method also includes, based on the plurality of slices, generating a set of candidate acoustic embeddings (232) where each candidate acoustic embedding includes a vector representation of acoustic features. The method further includes removing a subset of the candidate acoustic embeddings from the set of candidate acoustic embeddings. The method additionally includes generating an aggregate acoustic embedding (234) from the remaining candidate acoustic embeddings in the set of candidate acoustic embeddings after removing the subset of the candidate acoustic embeddings.