Speaker Embedding Networks With In-Network Audio Augmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional data augmentation methods for training neural networks in audio processing generate massive datasets, leading to resource-intensive training processes that are impractical for many entities due to resource limitations, affecting the quality and operation of neural networks.

Innovation Solution

Implementing in-network data augmentation layers within the neural network architecture that apply augmentation techniques directly on input audio signals, reducing the need for large datasets and minimizing resource consumption.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional data augmentation methods generate multiple copies of augmented data for each training speech sample, then the neural network training robustness is improved, but the dataset size becomes massive (terabytes) and resource consumption increases

Engineering Contradiction:
Improveneural network training robustnessVSAvoiddataset size
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent segments the data augmentation process from the dataset creation phase to the training phase. Instead of pre-generating all augmented copies into a massive dataset, the augmentation operations are divided and applied individually during neural network training through augmentation layers, reducing the need to store terabytes of augmented data.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies data augmentation operations as preliminary actions within the training pipeline itself. The augmentation layers perform transformations on-the-fly during training, so the robustness benefits are achieved through preliminary processing within the network rather than through pre-generated massive datasets.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If massive datasets containing multiple copies of augmented data are generated and stored, then the neural network training robustness is improved, but the computational resources and storage requirements become impractical for many entities

Engineering Contradiction:
Improveneural network training robustnessVSAvoidcomputational resource requirements
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent implements self-service data augmentation where the neural network itself performs the augmentation operations through dedicated augmentation layers during training. The system serves its own data needs by generating augmented samples on-demand within the training process, eliminating the need for external massive dataset generation and storage infrastructure.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent moves the data augmentation process from the data preparation dimension to the model architecture dimension. By embedding augmentation layers within the neural network, the system transforms how augmentation is applied - from a data-centric approach requiring massive storage to a model-centric approach that computes augmentations during training.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Adaptability or versatility

If multiple copies or segments of audio signals are generated for each augmentation operation, then the training data diversity is improved, but the processing time and computational load increase significantly

Engineering Contradiction:
Improvetraining data diversityVSAvoidprocessing time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent ensures continuous useful action by integrating data augmentation directly into the training flow through augmentation layers. Instead of discrete batch processing of augmented copies, the system continuously applies augmentation transformations during training iterations, maintaining data diversity without the overhead of generating and processing separate augmented datasets.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP4042415B1Z-vectors: speaker embeddings from raw audio using sincnet, extended CNN architecture, and in-network augmentation techniques
Publication Date: 2026.01.28 PINDROP SECURITY INC
  • EP4042415B1 patent drawingFigure 1
  • EP4042415B1 patent drawingFigure 2
  • EP4042415B1 patent drawingFigure 3

AI summary

Described herein are systems and methods for improved audio analysis using a computer-executed neural network having one or more in-network data augmentation layers. The systems described herein help ease or avoid unwanted strain on computing resources by employing the data augmentation techniques within the layers of the neural network. The in-network data augmentation layers will produce various types of simulated audio data when the computer applies the neural network on an inputted audio signal during a training phase, enrollment phase, and/or testing phase. Subsequent layers of the neural network (e.g., convolutional layer, pooling layer, data augmentation layer) ingest the simulated audio data and the inputted audio signal and perform various operations.