Electronic Device Voice Recognition with Wake-Word Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice recognition systems face challenges in balancing computational efficiency and accuracy, particularly in on-device environments, where reducing calculation and memory usage is crucial while maintaining high recognition rates, especially when ambient noise is present.
Innovation Solution
The system performs two-stage voice recognition: first recognizing a wake-up word using a reduced number of candidate texts and then recognizing instruction words using a higher number of candidate texts, optimizing resource usage and reducing errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If voice recognition is performed from the time point when wake-up starts, then voice recognition accuracy is improved, but calculation amount and memory usage increase
Solution Approach 1:
The patent segments the voice signal into two distinct sections: a first voice signal section corresponding to the wake-up word and a second voice signal section corresponding to the instruction word. Different voice recognition methods are applied to each section based on their specific characteristics, allowing optimized resource allocation for each segment rather than applying a uniform high-resource method to the entire signal.
Solution Approach 2:
The patent applies different voice recognition strategies to different parts of the voice signal. The first voice signal section (wake-up word) uses a method optimized for accuracy with more candidate texts, while the second voice signal section (instruction word) uses a method optimized for efficiency with fewer candidate texts, since the wake-up word has already been identified with high confidence.
2Quantity of substance
If voice recognition is performed from the time point when wake-up is completed, then calculation amount is reduced, but voice recognition rate decreases due to sensitivity to ambient noises
Solution Approach 1:
The patent divides the voice recognition process into two stages based on signal segmentation. The first stage processes the wake-up word section with enhanced accuracy to ensure reliable activation, while the second stage processes the instruction word section with reduced computational load, achieving a balance between accuracy and efficiency that neither extreme alone could achieve.
3Measurement precision
If wake-up word is recognized redundantly, then voice recognition accuracy is improved, but calculation amount and memory usage increase significantly
Solution Approach 1:
The patent performs preliminary voice recognition on the first voice signal section (wake-up word) to identify and confirm the wake-up word before proceeding to the second voice signal section (instruction word). This preliminary action ensures accurate wake-up detection while avoiding redundant processing of the instruction word, as the system already has high confidence in the wake-up word identification from the first stage.
Data Source
AI summary
A control method of an electronic device includes receiving a voice signal including a wake-up word and an instruction word for activating a voice recognition function, performing first voice recognition by acquiring at least one first candidate text in each frame of a first plurality of predetermined frames for a first voice signal section corresponding to the wake-up word in the voice, performing second voice recognition by acquiring at least one second candidate text in each frame of a second plurality of predetermined frames for a second voice signal section corresponding to the instruction word in the voice signal and performing second voice recognition, and providing a function corresponding to the instruction word based on results of the first voice recognition and the second voice recognition.