Wake Word-Free Command Fulfillment via Dual-Stage Audio Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated speech recognition (ASR) systems require wake words for user interaction, which can be inconvenient and may lead to incorrect rejections or false invocations, limiting user experience and privacy protection.
Innovation Solution
A system using a frame-level detector model and a word-level verifier model to detect target domain specific words, allowing for command fulfillment without a wake word, by training these models with self-supervised auxiliary tasks to identify commonly used commands and verify their presence in user utterances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a wake word detection model is implemented to protect user privacy, then cloud connection is established only when wake word is uttered, but the system requires continuous listening which increases power consumption and device resource usage
Solution Approach 1:
The audio processing is segmented into two distinct stages: frame-level detection for wake word identification and word-level verification for confirmation. This segmentation allows the system to quickly filter audio frames and only perform comprehensive verification on relevant segments, reducing overall power consumption while maintaining privacy protection
Solution Approach 2:
The system performs partial action by conducting frame-level detection on all audio input to identify potential wake word frames, then applies more resource-intensive word-level verification only to selected chunks. This partial verification approach reduces computational load and power consumption compared to full verification of all audio input
2Reliability
If wake word detection is implemented, then user privacy is protected by limiting cloud connection, but the system may incorrectly reject valid commands or falsely invoke the virtual assistant
Solution Approach 1:
The frame-level detector acts as an intermediary between audio input and the final wake word recognition decision. It pre-processes audio frames to identify candidate segments, which are then passed to the word-level verifier for final confirmation. This intermediary stage filters out obvious non-matches early while ensuring thorough verification of potential matches, improving overall detection accuracy
Solution Approach 2:
The system implements feedback through the two-stage verification process where frame-level detection results feed into word-level verification. The word-level verification outcomes provide feedback that can refine future frame-level detection, creating a closed-loop system that continuously improves wake word recognition accuracy while maintaining privacy protection
3Speed
If a lightweight frame-level detector model is used to reduce power consumption, then processing speed increases, but the model may lack the precision to accurately distinguish wake words from similar sounds
Solution Approach 1:
The detection process is segmented into frame-level detection using a lightweight model for rapid processing, followed by word-level verification using a more sophisticated model for accurate confirmation. This segmentation enables the system to maintain high processing speed in the first stage while achieving high accuracy in the second stage
Solution Approach 2:
The frame-level detector serves as an intermediary that performs rapid initial screening of audio frames. It identifies candidate frames that potentially contain wake words and passes them to the word-level verifier. This intermediary role allows the lightweight model to operate at high speed while the more accurate verification model handles the precision-critical task
Data Source
AI summary
A method comprises obtaining an audio input. The method also includes providing at least a portion of the audio input to a frame-level detector model. The method also includes obtaining a first output of the frame-level detector model including frame-level predictions associated with at least the portion of the audio input. The method also includes providing at least one chunked audio frame to a word-level verifier model. The method also includes obtaining a second output of the word-level verifier model including word-level probabilities associated with the at least one chunked audio frame. The method also includes instructing performance of automatic speech recognition on the audio input based on the word-level probabilities associated with the at least one chunked audio frame.