Secure Utterance Storage via Speech-to-Text De-identification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Voice-driven computing systems face challenges in securely storing utterances while efficiently utilizing computing resources, as storing recordings with high security levels requires significant processor cycles, memory, and storage space, and can derive personally identifiable information.

Innovation Solution

A method involving a speech-to-text service to translate and identify utterances, followed by a text-to-speech service to generate speech without personally identifiable information attributes, allowing the storage of this speech in a less secure location, thereby reducing resource usage.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If utterances are stored with high security levels to protect personally identifiable information, then privacy protection is improved, but computing resource utilization deteriorates

Engineering Contradiction:
Improveprivacy protectionVSAvoidcomputing resource utilization
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent extracts and removes personally identifiable information attributes from utterances using speech processing techniques. By separating the identifiable attributes from the core utterance content, the system can store the de-identified utterances in less secure (and thus more resource-efficient) storage locations while still protecting privacy. This directly resolves the contradiction by achieving privacy protection without requiring high-security storage resources.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the security level parameter of utterance storage by transforming the utterances from their original state (containing PII) to a de-identified state. This parameter change allows the same utterance data to be stored with lower security requirements, thereby reducing computing resource utilization while maintaining adequate privacy protection.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If original utterance recordings are stored for future use, then data availability is improved, but storage space requirements increase

Engineering Contradiction:
Improvedata availabilityVSAvoidstorage space
Core Design Contradiction:
ReliabilityVSVolume of stationary object

Solution Approach 1:

The patent creates de-identified copies of original utterances for storage. Instead of storing the original high-security utterances, the system generates copies with removed personally identifiable information attributes. These copies occupy storage space but enable future data availability without requiring the same level of security infrastructure, effectively reducing the storage burden while maintaining utility.

Inventive Principle:
Principle #26Copying

3Reliability

If speech-to-text and text-to-speech services are used to process utterances, then privacy protection is improved, but processing time increases

Engineering Contradiction:
Improveprivacy protectionVSAvoidprocessing time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent applies speech-to-text and text-to-speech processing as preliminary actions to remove personally identifiable information before storage. By performing this de-identification processing upfront, the system establishes privacy protection early in the data lifecycle. Although this initial processing takes time, it eliminates the need for ongoing security maintenance and enables faster subsequent operations on the de-identified data.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3646320B1Secure utterance storage
Publication Date: 2024.01.03 AMAZON TECH INC
  • EP3646320B1 patent drawingFigure 1
  • EP3646320B1 patent drawingFigure 2
  • EP3646320B1 patent drawingFigure 3

AI summary

Technologies for secure storage of utterances are disclosed. A computing device captures audio of a human making a verbal utterance. The utterance is provided to a speech-to-text (STT) service that translates the utterance to text. The STT service can also identify various speaker-specific attributes in the utterance. The text and attributes are provided to a text-to-speech (TTS) service that creates speech from the text and a subset of the attributes. The speech is stored in a data store that is less secure than that required for storing the original utterance. The original utterance can then be discarded. The STT service can also translate the speech generated by the TTS service to text. The text generated by the STT service from the speech and the text generated by the STT service from the original utterance are then compared. If the text does not match, the original utterance can be retained.