Delta Models for Privatized Speech-to-Text in Virtual Meetings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems in virtual meetings raise privacy concerns as they generate user-specific speech databases and models based on personally-identifiable information, lacking adequate control for users over their personal data.
Innovation Solution
The implementation of a delta model that generates and utilizes user-specific databases and models for speech recognition, while ensuring privacy by storing and running these models locally, and optionally in conjunction with a cloud-based model to reduce processing requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If user-specific speech databases and models are generated based on personally-identifiable information, then speech recognition accuracy is improved, but user privacy control deteriorates
Solution Approach 1:
The speech recognition system is segmented into two distinct models: a cloud-based model for general speech recognition and a local delta model for user-specific personalization. This segmentation allows user-specific data to be processed locally without being transmitted to the cloud, thereby maintaining speech recognition accuracy while protecting user privacy.
Solution Approach 2:
The delta model acts as an intermediary between the user's personal data and the speech recognition system. It processes user-specific information locally on the user's device, serving as a mediator that enables personalized speech recognition without directly exposing user data to external systems.
2Object-generated harmful factors
If user-specific models are stored and processed locally, then user privacy control is improved, but processing requirements and device complexity increase
Solution Approach 1:
The system implements local quality by placing only the essential user-specific delta model locally on the user's device, rather than storing complete speech recognition models. This approach provides privacy protection through local processing while minimizing the complexity and resource requirements on the user's device.
Solution Approach 2:
Instead of implementing complete speech recognition models locally, the system uses a partial approach by deploying only a delta model that contains user-specific adjustments. This partial action provides sufficient privacy protection while avoiding the excessive computational requirements of full local speech recognition.
3Power
If cloud-based speech recognition is used, then processing power is sufficient, but user data security deteriorates
Solution Approach 1:
The system extracts only the essential user-specific components (the delta model) and keeps them locally on the user's device, while leaving the computationally intensive general speech recognition model on the cloud. This extraction approach maintains data security by keeping sensitive user information local while still utilizing cloud processing power for general recognition tasks.
Data Source
AI summary
Provided herein are systems and methods for delta models for providing privatized speech-to-text during virtual meetings. In one embodiment, a system may include a non-transitory computer-readable medium; a communications interface; and a processor. The processor may be configured to execute processor-executable instructions to: join a virtual meeting. Each participant in the virtual meeting may exchange audio streams with other participants in the virtual meeting. The instructions may include receiving, from a video conference provider, a local model for speech recognition. The local model may be a copy of a centralized model. The instructions may include performing speech recognition using the local model on the audio streams. Performing speech recognition may include identifying audio feature data within the one or more audio streams, identifying, based on a vocabulary database, user-specific vocabulary within the audio feature data, and generating, based on the user-specific vocabulary, a private transcription of the audio streams.