Audio Watermark Detection for False Hotword Suppression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Unintentional activation of virtual assistants due to false hotword triggers in recorded speech leads to unnecessary energy consumption, network traffic, and potential safety risks, especially in noisy and reverberant environments.
Innovation Solution
An audio watermarking approach using a convolutional neural network-based detector to distinguish live speech from recorded speech, suppressing false hotword triggers by identifying audio watermarks in input data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If audio watermarking detection is implemented to distinguish live speech from recorded speech, then false hotword triggers are suppressed, but device complexity increases
Solution Approach 1:
An audio watermark detector is introduced as an intermediary component between the microphone input and the hotword recognition system. This detector analyzes the audio signal for presence of watermarks (indicating recorded speech) and suppresses hotword triggers when watermarks are detected, thereby resolving the contradiction by adding a mediating layer that improves reliability without fundamentally redesigning the entire system
Solution Approach 2:
The audio watermark detection is performed preliminarily before hotword recognition. By detecting watermarks in advance and determining whether the audio is recorded or live speech, the system can preemptively suppress false triggers, improving hotword trigger accuracy without requiring complex post-processing or reanalysis
2Reliability
If hotword detection is continuously monitored in recorded speech, then false triggers are detected, but energy consumption increases
Solution Approach 1:
The patent extracts and analyzes only specific audio features related to watermark presence rather than performing full hotword recognition on all recorded speech. By taking out only the essential detection function and separating it from the full speech processing pipeline, the system achieves false trigger detection with reduced energy consumption
Solution Approach 2:
The system performs partial action by implementing watermark detection without requiring complete hotword recognition processing. The audio watermark detector analyzes specific characteristics of the audio signal to identify watermarks, providing sufficient false trigger detection capability without the excessive energy cost of full speech processing
3Reliability
If audio watermarking is used to identify recorded speech, then false hotword triggers are suppressed, but processing time increases
Solution Approach 1:
The audio watermark detector focuses on analyzing specific local characteristics of the audio signal (watermark features) rather than processing the entire audio content. By concentrating computational resources on detecting specific watermark patterns in localized audio segments, the system achieves accurate recorded speech identification with minimal processing time
Solution Approach 2:
The system changes the detection parameter from comprehensive speech analysis to specific watermark feature detection. By transforming the problem from recognizing full speech content to detecting embedded watermark signatures, the processing time is significantly reduced while maintaining high reliability in identifying recorded speech
Data Source
AI summary
A method includes adding, by a first computing device, a first audio watermark to first speech data corresponding to playback of a first utterance including a hotword used to invoke an attention of a second computing device. The method includes outputting, by the first computing device, the playback of the first utterance corresponding to the watermarked first speech data. The second computing device is configured to receive the watermarked first speech data and determine to cease processing of the watermarked first speech data.