Keyphrase Detection Centerphone Selection for Low Resource Devices

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Keyphrase detection systems on low resource devices face challenges with high power consumption and inaccurate wake-ups due to inefficient memory usage and heavy computational loads, often triggering on words close to the actual keyphrase, leading to a poor user experience.

Innovation Solution

The implementation of a keyphrase detection system that uses a centerphone selection technique to reduce computational loads and memory requirements by selecting the most occurring center-phone triphones for the rejection model, combined with contextphone and silence modeling to enhance accuracy and reduce false wakes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional keyphrase detection systems are implemented on low resource devices, then detection capability is provided, but power consumption is high and accuracy is poor

Engineering Contradiction:

Improvekeyphrase detection accuracyVSAvoidpower consumption

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

The system segments the triphone set by identifying and selecting only the most occurring centerphones for the rejection model, separating the full triphone universe into manageable subsets that reduce computational load while maintaining detection accuracy

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The rejection model applies local quality optimization by using a specialized subset of triphones (centerphones) rather than the complete triphone set, tailoring the model's composition to the specific needs of wake word detection while reducing overall system resource requirements

Inventive Principle:

Principle #3Local quality

2Measurement precision

If comprehensive triphone models are used for accurate keyphrase detection, then detection accuracy improves, but memory usage and computational load increase

Engineering Contradiction:

Improvekeyphrase detection accuracyVSAvoidcomputational load

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The system extracts only the essential centerphone triphones from the complete triphone set for use in the rejection model, removing unnecessary complexity while retaining the critical information needed for accurate keyphrase detection

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The system dynamically selects triphones based on occurrence frequency, adapting the rejection model to use only the most relevant triphones for the specific wake word being detected, rather than maintaining a static comprehensive model

Inventive Principle:

Principle #15Dynamics

3Reliability

If traditional wake word detection is implemented, then basic functionality is provided, but false wake-ups occur frequently

Engineering Contradiction:

Improvefalse wake rateVSAvoidsystem simplicity

Core Design Contradiction:

ReliabilityVSEase of manufacture

Solution Approach 1:

The system performs preliminary selection of centerphones and construction of the rejection model before actual wake word detection, preparing the optimized model structure in advance to reduce false wake-ups without adding complexity to the real-time detection process

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS11127394B2Method and system of high accuracy keyphrase detection for low resource devices

Publication Date: 2021.09.21 INTEL CORP

AI summary

Techniques related to keyphrase detection for applications such as wake on voice are disclosed herein. Such techniques may have high accuracy by using scores of phone positions in triphones to select which triphones to use with a rejection model, using context-related phones for the rejection model, adding silence before keyphrase sounds for a keyphrase model, or any combination of these.