Speech Embedding Hashing for Secure Local User Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
User devices often lack the necessary computing resources to determine user identity from audio data in real-time, leading to delayed or insecure transmission of audio data to remote systems for processing, which may not be preferred by users due to privacy concerns and network availability issues.
Innovation Solution
The user device processes audio data using neural networks to determine user-specific embedding data, which is then transformed using a hash function to create second embedding data that can be sent to a remote system for user identification, allowing the device to determine user identity locally and securely without transmitting sensitive audio information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If audio data is transmitted to a remote system for speech processing, then speech recognition accuracy is improved, but network dependency and latency increase
Solution Approach 1:
The patent segments speech processing into two parts: local embedding extraction on the user device and remote user identification on the server. This allows the time-critical embedding extraction to occur locally without network latency, while the computationally intensive user identification occurs remotely, resolving the contradiction between accuracy and latency.
Solution Approach 2:
The patent extracts only the essential embedding data from audio signals locally, removing unnecessary audio data transmission. This extraction approach maintains speech processing accuracy while eliminating network transmission delays for the actual audio content, addressing both accuracy and latency requirements.
2Measurement precision
If audio data is transmitted to a remote system for user identification, then user identification accuracy is improved, but security and privacy risks increase
Solution Approach 1:
The patent extracts only the necessary embedding features from audio data locally before transmission, removing all other sensitive audio information. This extraction approach maintains user identification accuracy while minimizing privacy risks by transmitting only the essential identification data rather than the complete audio recording.
Solution Approach 2:
The patent introduces embedding data as an intermediary representation between the original audio data and user identification. This intermediary form preserves the identification information needed for accurate user recognition while eliminating personally identifiable audio characteristics, thus maintaining accuracy while reducing privacy risks.
3Productivity
If embedding data processing is performed locally on the user device, then processing speed is improved, but device computational requirements increase
Solution Approach 1:
The patent segments the computational workload by performing only embedding extraction locally on the user device, which is a relatively lightweight operation, while offloading the more computationally intensive user identification process to the remote server. This segmentation enables fast local processing without requiring the user device to have high computational resources.
Data Source
AI summary
A first neural network model of a user device processes audio data to extract audio embeddings that represent vocal characteristics of a user of an utterance represented in the audio data. The audio embeddings may then be hashed to remove characteristics specific to the user while still maintaining a unique set of characteristics. The hashed embeddings may be sent to a remote system, which may use them to identify the user.