Voice Trigger Accuracy via Dynamic Threshold Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice trigger systems for electronic devices often experience false acceptance and false rejection due to ambient noise, accent variations, and complexity in on-line adaptation or pre-training of key phrase voice models, leading to inefficiencies in waking up devices reliably.
Innovation Solution
The method involves extracting current and previous voice features, determining their similarity using algorithms like Viterbi or DTW, and adjusting confidence thresholds based on false acceptance or rejection situations to accurately wake up electronic devices, reducing the need for complex on-line adaptation or pre-training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If on-line adaptation or pre-training of key phrase voice model is performed, then the accuracy of voice trigger is improved, but the device complexity and cost increase
Solution Approach 1:
The patent extracts the voice feature comparison function from the complex on-line adaptation or pre-training process. Instead of performing comprehensive model adaptation, the system extracts and compares only the essential voice features (such as voice print, tone, rhythm) against stored reference features, significantly simplifying the complexity while maintaining accuracy.
Solution Approach 2:
The patent creates a simplified copy of the voice model by storing reference voice features of the user during initial setup. During subsequent operations, the system compares extracted voice features against this stored reference copy rather than performing complex real-time adaptation or training, reducing computational complexity while maintaining recognition accuracy.
2Measurement precision
If on-line adaptation or pre-training of key phrase voice model is performed, then the voice trigger accuracy is improved, but the time required for setup increases
Solution Approach 1:
The patent performs the voice feature extraction and reference storage as a preliminary action during initial user setup. Once the reference voice features are stored, the system can immediately use these pre-extracted features for comparison during normal operation, eliminating the need for time-consuming on-line adaptation or re-training during subsequent use.
3Object-generated harmful factors
If high confidence threshold is used, then false acceptance is reduced, but false rejection increases
Solution Approach 1:
The patent makes the confidence threshold dynamic rather than fixed. The system adjusts the threshold based on contextual factors such as ambient noise levels, user behavior patterns, and the degree of voice feature matching. This dynamic adjustment allows the system to reduce false acceptance when needed while maintaining high wake-up success rate through adaptive thresholding.
Solution Approach 2:
The patent implements feedback mechanisms where the system learns from user interactions and adjusts the confidence threshold accordingly. When users provide feedback about false acceptances or rejections, the system uses this information to refine the threshold setting, creating a closed-loop system that optimizes the balance between false acceptance reduction and wake-up reliability.
Data Source
AI summary
A voice trigger method for an electronic apparatus is provided. The voice trigger method includes the following steps. Receive a current voice signal. Extract a current voice feature from the current voice signal. Determine whether a previous voice feature of a previous voice signal is stored. When the previous voice feature is stored, at least one of adjusting a confidence threshold and determining whether to wake up the electronic apparatus is performed according to whether the previous voice feature is corresponding to a false acceptance situation or a false rejection situation and the similarity between the current voice feature and the previous voice feature.