Audio Front End Self-Trigger Prevention via Double-Talk Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Electronic devices face the challenge of self-triggering due to residual echo from output audio, which can incorrectly detect wakewords and lead to unwanted activation, especially during active playback when near-end speech is not present.
Innovation Solution
Implementing double-talk detection and suppression mechanisms within the Audio Front End (AFE) to differentiate between near-end, far-end, and double-talk conditions, using correlation data to determine when machine-generated audio is present and suppressing wakewords or audible words accordingly, thereby preventing self-triggering.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If echo cancellation is performed to remove residual echo from output audio, then the quality of audio processing is improved, but self-triggering may still occur when residual echo exceeds energy limits
Solution Approach 1:
The system performs double-talk detection before wakeword detection to preliminarily determine whether near-end speech is present. By proactively identifying playback states and generating DTD data in advance, the system prevents self-triggering before it occurs, rather than merely attempting to cancel echo after the fact.
Solution Approach 2:
The patent introduces DTD data as an intermediary signal that mediates between the audio processing pipeline and wakeword detection. This intermediary carries information about near-end speech presence and controls whether wakeword detection should proceed, effectively filtering out false triggers from residual echo.
2Speed
If wakeword detection is continuously monitored to detect voice commands, then voice command responsiveness is improved, but false activation from machine-generated audio increases
Solution Approach 1:
The system dynamically adjusts wakeword detection behavior based on real-time DTD data. During playback periods when near-end speech is absent, the system suppresses wakeword detection to prevent false activations. When near-end speech is detected, the system immediately resumes normal wakeword detection, maintaining responsiveness when needed.
Solution Approach 2:
The patent changes the operational parameters of wakeword detection based on playback state. By monitoring energy levels and correlation values in real-time, the system adjusts detection sensitivity and activation thresholds dynamically, reducing false positives during machine-generated audio playback while maintaining high sensitivity for genuine voice commands.
3Reliability
If double-talk detection is implemented to differentiate between near-end and far-end speech, then self-trigger prevention is improved, but device complexity increases
Solution Approach 1:
The AFE component performs multiple functions within a single integrated module: echo cancellation, double-talk detection, correlation analysis, and wakeword detection control. By consolidating these functions into one universal component rather than separate modules, the system achieves high self-trigger prevention accuracy without proportionally increasing overall device complexity.
Solution Approach 2:
The system uses its own existing audio processing resources to perform double-talk detection. The AFE leverages the same microphones, signal processing pipelines, and computational resources already dedicated to echo cancellation and wakeword detection, making the DTD functionality self-contained and avoiding the need for additional dedicated hardware or complex external systems.
Data Source
AI summary
A system configured to perform self-trigger prevention to avoid a device waking itself up when a wakeword is output by the device's own output audio. For example, during active playback the device may perform double-talk detection and suppress wakewords or other device-directed utterances when near-end speech is not present. To detect whether near-end speech is present, an Audio Front End (AFE) of the device may perform echo cancellation and generate correlation data indicating an amount of correlation between an output of the echo canceller and an estimated reference signal. When the correlation is high in certain frequency ranges, near-end speech is not present and the device may suppress the utterance. When the correlation is low, indicating that near-end speech could be present, the device does not suppress the utterance and sends the utterance to a remote system for speech processing.