Default Assistant Fallback Latency Reduction in Multi-Assistant Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Multi-assistant systems experience latency issues when determining whether an utterance includes a wakeword, leading to delayed processing and user frustration, as they typically wait until the end of the utterance to select the appropriate speech-processing system.
Innovation Solution
Implementing a timer-based mechanism that starts upon detecting voice activity, allowing for early determination of wakeword absence by using voice activity detection (VAD) and timing information for wakewords, enabling the default speech-processing system to begin processing the utterance sooner, potentially within a second of the beginning of speech.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the system waits until the end of the utterance to determine wakeword presence, then wakeword detection accuracy is improved, but user-perceived latency increases
Solution Approach 1:
The system performs preliminary voice activity detection and timing analysis during the utterance before final wakeword determination. By starting a timer upon detecting voice activity and comparing utterance duration against expected wakeword durations, the system preliminarily assesses wakeword presence without waiting for complete utterance processing, thereby reducing perceived latency while maintaining accuracy through subsequent confirmation.
2Adaptability or versatility
If the system processes multiple assistants simultaneously, then system versatility is improved, but processing complexity increases
Solution Approach 1:
The system segments the multi-assistant processing into distinct phases: voice activity detection phase, timing analysis phase, and final routing phase. Each assistant's wakeword detection operates independently with its own timer, allowing parallel processing without full system complexity. This segmentation enables versatile multi-assistant support while managing processing complexity through structured separation of detection and decision-making functions.
Data Source
AI summary
A speech-processing system may provide access to one or more virtual assistants via an audio-controlled device. A virtual assistant may be invoked by speaking a wakeword. In some cases, a default virtual assistant may be invoked when an utterance is spoken without a preceding wakeword. Such a multi-assistant speech-processing system may make an early determination that a received utterance does not include a wakeword, and begin processing the utterance prior to completion of the utterance, thereby reducing user perceived latency. For example, the system may start a timer when a gesture and/or voice activity is detected. If no wakeword is detected after a time corresponding to speaking durations of known wakewords, the system may determine that the utterance is to be processed according to the default virtual assistant.