Secondary Wakeword Detector for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition systems face issues with bandwidth consumption and privacy concerns due to continuous audio transmission, and they often waste resources processing non-command audio, while inadvertently detecting wakewords in output audio can lead to unintended device activation.
Innovation Solution
Implementing a local device configuration to only activate upon a user-speaking wakeword, using a secondary wakeword detector to ignore inadvertent wakewords in output audio by disabling primary wakeword detection during output, and employing techniques like acoustic echo cancellation and audio fingerprinting to differentiate between user-generated and device-generated wakewords.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If continuous audio transmission is used for speech recognition, then speech processing capability is improved, but bandwidth consumption increases
Solution Approach 1:
The system performs preliminary wakeword detection locally before transmitting audio to the server. By detecting the wakeword in advance and only then initiating audio transmission, the system avoids continuous bandwidth consumption while maintaining speech processing capability when needed.
2Reliability
If continuous audio transmission is used for speech recognition, then speech processing capability is improved, but user privacy is compromised
Solution Approach 1:
The system performs preliminary wakeword detection locally on the device before transmitting audio to the server. This preliminary action filters out non-command audio locally, so only relevant audio segments are transmitted, thereby preserving user privacy while maintaining speech processing capability.
3Measurement precision
If resources are allocated for processing all audio, then speech recognition accuracy is improved, but resource waste increases
Solution Approach 1:
The system performs preliminary wakeword detection to identify when a command is being spoken. Only after detecting the wakeword does the system allocate full resources for speech recognition processing. This preliminary filtering action ensures high recognition accuracy for actual commands while avoiding resource waste on non-command audio.
4Speed
If primary wakeword detection is always active, then device responsiveness is improved, but unintended activation increases
Solution Approach 1:
The system segments wakeword detection into two independent components: a primary detector that provides fast initial detection and a secondary detector that provides verification. The primary detector remains always active for responsiveness, while the secondary detector verifies wakewords to prevent unintended activation, thus resolving the contradiction between speed and reliability.
Solution Approach 2:
The secondary wakeword detector acts as an intermediary verification layer between the primary detector and device activation. When the primary detector identifies a potential wakeword, the secondary detector verifies it before triggering device activation. This intermediary mechanism maintains fast responsiveness while ensuring activation accuracy.
Data Source
AI summary
A system and method for temporarily disabling keyword detection to avoid detection of machine-generated keywords. Audio data received for output by audio speakers is first captured by an effect component, such as an audio equalizer. The effect component may perform various operations including altering the audio data, copying the audio data and delaying the time in which the output audio is sent to the audio speakers. The effect component may generate a copy of the audio data and transmit or route the copy of the audio data to a secondary keyword detector via an audio channel, resulting in a further delay. The secondary detector may determine that the copy of the processed audio data includes a keyword that is likely to be output during a first time interval. The secondary detector may then transmit a signal to a primary keyword detector to disable keyword detection during the first time interval of the output audio.