Speech Emotion Estimation Using Disentangled Vector Separation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing emotion estimation technologies, such as those disclosed in JP 2019-28910 A, do not effectively utilize machine learning to estimate customer emotions during business negotiations, limiting their application in improving customer service and support.
Innovation Solution
An emotion estimation method that separates speech data into first and second vector data, utilizing a learning model trained with loss functions based on linguistic and symmetric/asymmetric learning, and mutual information minimization to estimate expressive and intrinsic emotions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional dialogue analysis systems are used to analyze business negotiation contents, then communication compliance can be checked, but emotion estimation accuracy is insufficient
Solution Approach 1:
The speech data is segmented into multiple feature dimensions including linguistic information, acoustic features, and prosodic features. The learning model is divided into multiple processing branches that handle different feature types separately before integrating them for comprehensive emotion estimation, thereby improving accuracy without overwhelming system complexity
Solution Approach 2:
A feature extraction module serves as an intermediary between raw speech data and the emotion estimation model. This intermediary processes and transforms speech signals into meaningful feature representations, enabling accurate emotion detection while keeping the main estimation model relatively simple and manageable
2Measurement precision
If speech data is processed using simple analysis methods, then processing speed is maintained, but emotion recognition precision is insufficient
Solution Approach 1:
Speech data undergoes preliminary feature extraction and preprocessing before being fed into the emotion estimation model. Acoustic features, prosodic features, and linguistic information are extracted and organized in advance, which reduces the computational burden during actual emotion estimation and maintains efficient processing speed while achieving high precision
Solution Approach 2:
The system dynamically adjusts processing based on input characteristics. The learning model selectively processes different feature combinations depending on the speech data properties, optimizing the balance between processing speed and emotion recognition precision for different types of input
3Measurement precision
If comprehensive speech features are analyzed, then emotion estimation accuracy improves, but computational complexity increases
Solution Approach 1:
Comprehensive speech features are segmented into distinct categories (linguistic information, acoustic features, prosodic features) that are processed through specialized sub-modules. This segmentation allows the system to handle complex features in an organized manner, improving emotion estimation accuracy while managing computational complexity through modular architecture
Solution Approach 2:
The learning model is designed with multi-functionality to handle various feature types uniformly. A single integrated model processes multiple feature dimensions simultaneously, reducing the need for separate specialized models and thereby controlling overall computational complexity while maintaining high estimation accuracy
Data Source
AI summary
An emotion estimation method executed by an information processing device, comprising: acquiring speech data; inputting speech data into a learning model; separating the speech data into at least first vector data and second vector data; and estimating an emotion corresponding to the speech data based at least on the first vector data and the second vector data, wherein the learning model is trained based on a first loss function based on a difference between the linguistic information based on the speech data and the first vector data, a second loss function based on symmetric learning or asymmetric learning of the second vector data, and a third loss function that minimizes a mutual information amount between the first vector data and the second vector data.