Privacy-Preserving Speech Recognition Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current speech recognition systems face challenges in handling sensitive data, such as contact information and location data, which may need to be deleted quickly due to privacy restrictions, and require substantial computational resources for training, making it difficult to achieve high-quality recognition in privacy-sensitive environments.

Innovation Solution

The system identifies sensitive data during speech recognition, uses it to train an automated speech recognizer before deletion, and incorporates distributed backpropagation neural network training, allowing continuous model adaptation with real-time data points, enabling efficient operation in environments with strict logging requirements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If sensitive data is stored in logs for training, then model training quality improves, but privacy compliance deteriorates due to data retention requirements

Engineering Contradiction:
Improvemodel training qualityVSAvoidprivacy compliance
Core Design Contradiction:
Manufacturing precisionVSObject-affected harmful factors

Solution Approach 1:

The system performs preliminary actions by identifying sensitive data and deleting it before the data can cause privacy compliance issues. The deletion is scheduled to occur after a predetermined period, ensuring the data is removed timely while still allowing sufficient training benefit.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system extracts only the necessary training information from the audio data while leaving behind sensitive information that needs to be deleted. By separating the useful training signals from sensitive data, the system can train models effectively without retaining privacy-sensitive information in logs.

Inventive Principle:
Principle #2Taking out (Extraction)

2Object-affected harmful factors

If data is deleted quickly to meet privacy requirements, then privacy compliance improves, but model training effectiveness deteriorates

Engineering Contradiction:
Improveprivacy complianceVSAvoidmodel training effectiveness
Core Design Contradiction:
Object-affected harmful factorsVSManufacturing precision

Solution Approach 1:

The system performs preliminary identification and scheduling of data deletion before the data expires. By planning the deletion timing in advance, the system ensures privacy compliance while maximizing the training window for model effectiveness.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system uses feedback mechanisms to monitor data sensitivity and adjust training processes accordingly. By continuously assessing data characteristics and deletion requirements, the system optimizes the balance between privacy compliance and training effectiveness.

Inventive Principle:
Principle #23Feedback

3Use of energy by moving object

If distributed backpropagation training is used, then computational resource requirements improve, but system complexity increases

Engineering Contradiction:
Improvecomputational resource efficiencyVSAvoidsystem complexity
Core Design Contradiction:
Use of energy by moving objectVSDevice complexity

Solution Approach 1:

The training system is segmented into distributed components that can process data independently across multiple nodes. By dividing the training workload into smaller segments handled by different computational units, the system reduces the burden on any single resource while maintaining overall training effectiveness.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system introduces intermediary components that facilitate communication and coordination between distributed training nodes. These intermediaries manage the complexity of distributed backpropagation by handling data shuffling, gradient aggregation, and synchronization tasks automatically.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3234944B1Privacy-preserving training corpus selection
Publication Date: 2023.02.22 GOOGLE LLC
  • EP3234944B1 patent drawingFigure 1
  • EP3234944B1 patent drawingFigure 2
  • EP3234944B1 patent drawingFigure 3

AI summary

The present disclosure relates to training a speech recognition system. A system that includes an automated speech recognizer and receives data from a client device. The system determines that at least a portion of the received data is likely sensitive data. Before the at least a portion of the received data is deleted, the system provides the at least a portion of the received data to a model training engine that trains recognition models for the automated speech recognizer. After the at least a portion of the received data is provided, the system deletes the at least a portion of the received data.