Multi-Wake Phrase Detection for Low-Power Voice Assistants
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current voice activation systems in battery-powered devices consume high power due to continuous running of multiple vendor-specific wake phrase detection algorithms, leading to increased MIPS and clock speed, especially when multiple voice assistants are activated simultaneously.
Innovation Solution
Implementing a multi-wake phrase detection system that uses a single Voice Activity Detection process to identify multiple wake phrases with a low miss-rate, activating only the necessary vendor-specific algorithms and reducing parallel processing, thereby minimizing power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple vendor-specific wake phrase detection algorithms are run in parallel, then the wake phrase detection coverage is improved, but the power consumption increases
Solution Approach 1:
The system segments the wake phrase detection process into two stages: first a universal VAD process detects speech presence, then vendor-specific algorithms are selectively activated only for detected speech segments. This segmentation allows multiple wake phrase detectors to operate without continuously processing all audio streams through all algorithms, reducing overall power consumption while maintaining comprehensive detection coverage.
Solution Approach 2:
Instead of running all vendor-specific wake phrase algorithms continuously or in parallel for every audio input, the system applies partial action by activating only the necessary algorithms based on VAD detection results. This selective activation ensures that computational resources are expended only when and where needed, reducing power consumption while maintaining adequate detection coverage for multiple voice assistants.
2Measurement precision
If the on-device processor is activated to run multiple wake phrase algorithms, then the wake phrase detection accuracy is improved, but the MIPS and clock speed requirements increase
Solution Approach 1:
The VAD process performs preliminary action by pre-processing the audio stream and identifying segments containing human speech before activating vendor-specific wake phrase algorithms. This preliminary filtering ensures that complex wake phrase detection algorithms are only executed when speech is actually present, reducing the overall computational burden on the on-device processor while maintaining high detection accuracy for wake phrases.
Solution Approach 2:
The VAD process acts as an intermediary between the audio input and vendor-specific wake phrase algorithms. It mediates by filtering and preparing audio segments, passing only relevant speech-containing segments to the wake phrase detectors. This intermediary role reduces the computational complexity required for wake phrase detection by eliminating the need to process silent or non-speech audio segments through multiple algorithms.
3Reliability
If all vendor-specific wake phrase algorithms are executed for each VAD detection, then the wake phrase detection reliability is improved, but the power consumption increases
Solution Approach 1:
The system dynamically adjusts the execution of wake phrase algorithms based on real-time VAD detection results. Instead of statically running all algorithms for every audio segment, the system dynamically activates only those algorithms necessary for the current speech detection context. This dynamic approach maintains reliable wake phrase detection by ensuring appropriate algorithms are run when speech is present while reducing power consumption by avoiding unnecessary algorithm execution during silent periods.
Data Source
AI summary
Systems and methods presented herein generally include multi-wake phrase detection executed on a single device utilizing multiple voice assistants. Systems and methods presented herein can further include continuously running a Voice Activity Detection (VAD) process which detects presence of human speech. The multi-wake phrase detection can activate when the VAD process detects human speech. Once activated, the multi-wake phrase detection can determine which (if any) of the wake phrases of the multiple voice assistants might be in the detected speech. Operation of the multi-wake phrase detection can have a low miss-rate. In some examples, operation of the multi-wake phrase detection can be granular to accomplish the low miss-rates at low power with a tolerance for false positives on wake phrase detection.