Hybrid Speech Recognition System for Privacy and Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Local automatic speech recognition systems in vehicles or mobile devices lack the computing power to match network-based systems, resulting in inferior recognition performance, and there is a need to utilize private user data for improved recognition without sharing it with network-based systems.
Innovation Solution
A hybrid recognition system that combines an embedded local speech recognition system with a network-based system, where the embedded recognizer accesses private user data to enhance speech recognition locally while minimizing network bandwidth usage and maintaining data privacy, using a hybrid approach that allows both recognizers to work serially or in parallel to achieve accurate results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a local speech recognition system is used in a vehicle or mobile device, then data privacy is maintained and network bandwidth usage is reduced, but recognition accuracy and computing power are insufficient compared to network-based systems
Solution Approach 1:
The speech recognition system is divided into two segments: a local embedded recognizer that handles privacy-sensitive tasks and basic recognition, and a remote network-based recognizer that handles complex recognition tasks requiring higher computing power. This segmentation allows each component to operate within its capability limits while achieving overall system goals.
Solution Approach 2:
A hybrid recognition system acts as an intermediary between the local embedded recognizer and the remote network-based recognizer. It coordinates between the two systems, determining when to use local processing (for privacy and speed) and when to engage the remote system (for accuracy and complexity), thus resolving the contradiction between local constraints and remote capabilities.
2Reliability
If private user data is shared with a network-based speech recognition system, then recognition accuracy is improved, but data privacy is compromised
Solution Approach 1:
The system extracts and processes only the necessary portions of private user data locally before sending anonymized or processed information to the network-based system. This extraction approach allows the network system to improve accuracy without accessing complete private data sets, thus protecting user privacy while still benefiting from enhanced recognition.
Solution Approach 2:
The hybrid recognition system serves as an intermediary that mediates between local privacy concerns and network-based accuracy requirements. It selectively shares only non-sensitive information with the remote system while keeping sensitive private data localized, thus resolving the privacy-accuracy trade-off.
3Reliability
If a network-based speech recognition system is used, then recognition accuracy is improved, but network bandwidth usage increases and connectivity dependency increases
Solution Approach 1:
The system segments speech recognition tasks based on their complexity and privacy requirements. Simple, routine recognition tasks are handled locally by the embedded system, reducing network bandwidth usage. Only complex tasks requiring superior accuracy are transmitted to the network-based system, optimizing bandwidth utilization while maintaining accuracy where needed.
Solution Approach 2:
The local embedded speech recognition system provides self-service for basic recognition tasks, handling them autonomously without network intervention. This self-service capability reduces dependency on network connectivity and minimizes bandwidth consumption, while the system can still leverage network resources when enhanced accuracy is required.
Data Source
AI summary
A method of providing hybrid speech recognition between a local embedded speech recognition system and a remote speech recognition system relates to receiving speech from a user at a device communicating with a remote speech recognition system. The system recognizes a first part of speech by performing a first recognition of the first part of the speech with the embedded speech recognition system that accesses private user data, wherein the private user data is not available to the remote speech recognition system. The system recognizes the second part of the speech by performing a second recognition of the second part of the speech with the remote speech recognition system. The final recognition result is a combination of these two recognition processes. The private data can be such local information as a user location, a playlist, frequently dialed numbers or texted people, user contact list information, and so forth.