Dual Speech Recognizer Segmentation for Privacy and Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition systems face challenges in accurately recognizing terms and names that are not part of their general vocabulary, which can lead to privacy concerns and reduced recognition accuracy for user-specific phrases.
Innovation Solution
A dual-speech recognition system comprising a limited speech recognizer and an expanded speech recognizer, where the limited recognizer focuses on user-specific contact names and voice commands, while the expanded recognizer handles general grammar and proper names, with a transcription aligner and disambiguator combining their outputs to enhance recognition and privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a general-purpose speech recognition system is used, then it can recognize common words and phrases, but it fails to accurately recognize user-specific terms and contact names
Solution Approach 1:
The speech recognition system is divided into two separate recognizers: a limited speech recognizer with a focused vocabulary for user-specific terms and contact names, and an expanded speech recognizer with a comprehensive vocabulary for general speech. This segmentation allows each recognizer to excel at its specialized domain, resolving the contradiction between recognition accuracy for specific terms and vocabulary coverage.
Solution Approach 2:
A transcription aligner and disambiguator act as intermediary components that combine the outputs of both recognizers. The aligner matches transcriptions from both recognizers, and the disambiguator resolves conflicts by determining which recognizer's output is more reliable in each context. This intermediary mechanism enables the system to leverage both the specificity of the limited recognizer and the comprehensiveness of the expanded recognizer.
2Measurement precision
If user-specific contact information is processed by a network-based recognizer, then recognition accuracy improves, but privacy concerns increase
Solution Approach 1:
The system segments the speech recognition task by routing user-specific contact information through a local limited speech recognizer that processes data device-side, while general speech is handled by the network-based expanded recognizer. This segmentation keeps sensitive information local, reducing privacy risk while maintaining recognition accuracy for contact names through the specialized local model.
3Adaptability or versatility
If a single comprehensive speech recognizer is used, then vocabulary coverage is maximized, but the system becomes less efficient and more complex
Solution Approach 1:
Instead of using a single comprehensive recognizer, the system segments recognition into two specialized models with different vocabulary scopes. The limited recognizer handles user-specific terms efficiently, while the expanded recognizer handles general speech. This segmentation reduces the computational burden on each individual model and allows for more efficient processing.
Solution Approach 2:
The limited speech recognizer uses a focused, minimal vocabulary specifically tailored for user-specific terms and contact names, rather than maintaining a complete general-purpose vocabulary. This partial action approach allows the local model to be more efficient and accurate for its specific domain while the expanded recognizer supplements it for general speech, reducing overall system complexity.
Data Source
Figure 1
Figure 2A~2C
Figure 2D~2F
AI summary
The subject matter of this specification can be embodied in, among other things, a method that includes receiving audio data that corresponds to an utterance, obtaining a first transcription of the utterance that was generated using a limited speech recognizer. The limited speech recognizer includes a speech recognizer that includes a language model that is trained over a limited speech recognition vocabulary that includes one or more terms from a voice command grammar, but that includes fewer than all terms of an expanded grammar. A second transcription of the utterance is obtained that was generated using an expanded speech recognizer. The expanded speech recognizer includes a speech recognizer that includes a language model that is trained over an expanded speech recognition vocabulary that includes all of the terms of the expanded grammar. The utterance is classified based at least on a portion of the first transcription or the second transcription.