Streaming Encrypted Content Translation With Local Speech Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing translation technologies rely on datasets comprising actual user signals, disregarding data privacy and requiring complete message reception before translation, which hinders live conversations across different languages.
Innovation Solution
A system where user devices communicate via end-to-end encrypted links, locally training speech models using encrypted user inputs, and employing federated learning to generate and update models without exposing actual conversations to the server.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If translation services are provided using actual user signals acquired by a server, then translation accuracy can be improved, but user data privacy is compromised
Solution Approach 1:
The patent introduces an intermediary mechanism where the server provides training data to user devices in an encrypted or anonymized form, rather than directly collecting and processing actual user signals. This intermediary layer allows the system to improve translation models through training while protecting user privacy by preventing direct access to raw personal data.
Solution Approach 2:
The system creates synthetic or anonymized copies of training data that preserve the statistical properties needed for model training without containing actual sensitive user information. These copied datasets can be freely shared and processed without compromising the privacy of original users.
2Measurement precision
If translation is performed only after complete message reception, then translation accuracy is improved, but real-time communication capability deteriorates
Solution Approach 1:
The system performs translation on partial messages as they arrive, rather than waiting for complete messages. This partial action approach enables real-time translation during ongoing conversations, accepting that some translation may be performed on incomplete data rather than delaying until full messages are received.
Solution Approach 2:
The system performs preliminary translation on incoming message segments as they arrive, preparing translations in advance before the complete message is received. This preliminary action allows the translation process to overlap with message transmission, reducing overall communication delay.
Data Source
AI summary
Method and servers for generating a speech model for generating signals representative of utterances in a first language based on signals representative of utterances in a second language are disclosed. The method comprises transmitting a first and a second speech models to a first and a second devices of a first and a second users respectively. The first device is communicatively coupled with the second device by an encrypted communication link. A third speech model is acquired from the second device based on a local training of the second speech model on the second device. A training set comprises a first and a second decrypted signals representative of an utterance of the first user in the first language and a translated utterance of the first user in the second language respectively. The speech model is locally generated by the server by combining the second and third speech models.