Adaptive Voice Interaction Model Selection for Response Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing voice interaction systems face challenges in preventing response errors due to the use of inappropriate learning models, which can lead to overlapping user and system speech or long silences, as they rely on a single machine learning model for all user interactions regardless of the user's emotional or active state.

Innovation Solution

A voice interaction system that includes a user state detection unit to identify the user's state, such as activeness, emotion, or health condition, and a learning model selection unit to switch to a suitable learning model based on the detected state, ensuring accurate responses tailored to the user's current condition.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a single machine learning model is used for all user interactions, then the system structure is simple, but response accuracy deteriorates when the user state changes

Engineering Contradiction:
Improvesystem structureVSAvoidresponse accuracy
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The patent applies the dynamics principle by making the learning model selection adaptive and dynamic based on user state. Instead of using a fixed single model, the system dynamically selects from multiple learning models according to the detected user state (emotional state, activity level, etc.), allowing the system to adapt its response determination mechanism to changing user conditions while maintaining manageable complexity through automated state-based selection.

Inventive Principle:
Principle #15Dynamics

2Reliability

If multiple learning models are prepared for different user states, then response accuracy improves, but device complexity increases

Engineering Contradiction:
Improveresponse accuracyVSAvoidsystem structure
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies segmentation by dividing the response determination function into multiple specialized learning models, each optimized for specific user states. The system segments the user state space (emotional states, activity levels) and assigns appropriate learning models to each segment, allowing high response accuracy for each state while managing complexity through modular model organization and automated selection mechanisms.

Inventive Principle:
Principle #1Segmentation

3Ease of operation

If the learning model is fixed, then the system is easy to operate, but response errors occur when user state changes

Engineering Contradiction:
Improvesystem operationVSAvoidresponse accuracy
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The patent applies self-service by implementing automated user state detection and automatic learning model selection based on detected states. The system independently monitors user state changes and autonomously switches between learning models without requiring manual intervention, maintaining ease of operation while improving response accuracy through adaptive model selection that responds to changing user conditions.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3618063B1Voice interaction system, voice interaction method and corresponding program
Publication Date: 2022.09.14 TOYOTA JIDOSHA KK
  • EP3618063B1 patent drawingFigure 1
  • EP3618063B1 patent drawingFigure 2
  • EP3618063B1 patent drawingFigure 3

AI summary

A voice interaction system capable of appropriately handling a situation so as to effectively prevent a response error from occurring is provided. A speech acquisition unit 102 acquires user speech. A feature extraction unit 104 extracts a feature of the user speech. A response determination unit 120 determines a response corresponding to the extracted feature vector using any one of a plurality of learning models. A response execution unit 130 executes the determined response. A user state detection unit 140 detects a user state. A learning model selection unit 150 selects a learning model from a plurality of learning models in accordance with the detected user state. The response determination unit 120 determines a response using the selected learning model.