Always-Listening Voice Detection With Thresholded Spectral Wake-Up
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice-controlled consumer electronic devices face challenges in reducing average current consumption while maintaining low voice-command misdetection rates, requiring an efficient always-listening mode.
Innovation Solution
A low-power, always-listening voice-command detection and capture system with discrete operational states, utilizing a voice-activity detector (VAD) that compares digital-audio signal amplitudes to thresholds, performs spectral analysis, and generates interrupt signals to reduce current consumption and misdetection rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the device operates in always-listening mode with high sensitivity to detect voice commands, then the voice-command detection accuracy is improved, but the average current consumption increases
Solution Approach 1:
The voice detection system is segmented into multiple operational states (low-power state and high-power state) with different detection sensitivities. The VAD operates in these segmented states based on signal conditions, allowing the system to achieve high detection accuracy when needed while consuming low power during normal operation.
Solution Approach 2:
The system dynamically transitions between operational states based on real-time signal analysis. When the VAD detects signal characteristics indicating a voice command, the system transitions from low-power state to high-power state, enabling adaptive detection accuracy that responds to actual voice presence rather than maintaining constant high sensitivity.
2Use of energy by moving object
If the device reduces current consumption by operating in low-power state, then the average current consumption is reduced, but the voice-command misdetection rate increases
Solution Approach 1:
The system performs preliminary voice activity detection in the low-power state by analyzing signal amplitude and spectral characteristics. This preliminary action identifies potential voice commands before full processing begins, allowing the system to wake from low-power state in advance of actual voice commands and prevent misdetections.
Solution Approach 2:
The system uses feedback from spectral analysis and amplitude detection to adjust its operational state. When the VAD detects signal characteristics consistent with voice activity, it triggers a state transition to high-power mode, using the feedback from low-power detection to prevent misdetections while maintaining overall low power consumption.
3Reliability
If the VAD performs spectral analysis continuously to improve voice detection accuracy, then the voice-command detection accuracy is improved, but the processing time and power consumption increase
Solution Approach 1:
Spectral analysis is performed periodically rather than continuously. The VAD monitors signal amplitude continuously in low-power state and only initiates spectral analysis when amplitude thresholds indicate potential voice activity. This periodic execution of computationally intensive spectral analysis reduces processing time and power consumption while maintaining detection accuracy.
Data Source
AI summary
A system for detecting and capturing voice commands, the system comprising a voice-activity detector (VAD) configured to receive a VAD-received digital-audio signal; determine the amplitude of the VAD-received digital-audio signal; compare the amplitude of the VAD-received digital-audio signal to a first threshold and to a second threshold; withhold a VAD interrupt signal when the amplitude of the VAD-received digital-audio signal does not exceed the first threshold or the second threshold; generate the VAD interrupt signal when the amplitude of the VAD-received digital-audio signal exceeds the first threshold and the second threshold; and perform spectral analysis of the VAD-received digital-audio signal when the amplitude of the VAD-received digital-audio signal is between the first threshold and the second threshold.