Audio Signal Augmentation for Cost-Efficient Speaker Recognition Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The collection of training data for speaker recognition is costly and limited by the number of individuals from whom data is collected, hindering the improvement of feature extraction accuracy in existing systems.

Innovation Solution

An audio processing apparatus and method that generate training data by signal processing existing audio signals to imitate different voice pitches and vocal tract lengths, reducing the need for extensive data collection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If training data is collected from many people to improve feature extraction accuracy, then the extraction accuracy of the feature extractor is improved, but the cost of collecting training data increases significantly

Engineering Contradiction:
Improveextraction accuracyVSAvoidcost of collecting training data
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent generates synthetic training data by copying and transforming existing audio signals through signal processing techniques. Instead of collecting original audio data from many individuals, the system creates artificial training samples by applying transformations to available audio data, thereby reducing data collection costs while maintaining sufficient training data quantity and diversity.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent applies signal processing transformations that modify parameters of existing audio signals, such as pitch shifting, time stretching, and noise addition. These parameter changes create diverse training variations from limited source data, enabling the feature extractor to learn robust features without requiring extensive data collection from numerous individuals.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If the number of people from whom training data is collected is increased to improve speaker recognition performance, then the reliability of speaker recognition is improved, but the complexity and cost of the data collection process increases

Engineering Contradiction:
Improvespeaker recognition performanceVSAvoiddata collection process complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system generates synthetic training data by copying and transforming existing audio recordings. This approach eliminates the need to coordinate data collection from numerous individuals, significantly simplifying the data collection process while providing sufficient training data to improve speaker recognition reliability through enhanced feature extraction.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system performs self-service by generating its own training data from available audio recordings through automated signal processing. This eliminates the need for external data collection efforts involving multiple participants, reducing process complexity while maintaining the ability to improve recognition performance through data augmentation.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3872808B1Audio processing apparatus, audio processing method, and program
Publication Date: 2025.12.24 NEC CORP
  • EP3872808B1 patent drawingFigure 1
  • EP3872808B1 patent drawingFigure 2
  • EP3872808B1 patent drawingFigure 3

AI summary

An audio processing apparatus 100 is apparatus for generating a training data in speaker recognition. The audio processing apparatus 100 includes a data acquisition unit configured to acquire an audio signal that is a source of the training data as sample data, a data generation unit configured to executes signal processing on the acquired sample data, and to generates a new audio signal as the training data whose similarity with the sample data is within the set range.