Method for emotion recognition based on minimum classification error

• US 8,180,638 B2
• Filed: 02/23/2010
• Issued: 05/15/2012
• Est. Priority Date: 02/24/2009
• Status: Expired due to Fees

• Current Assignee: Electronics and Telecommunications Research Institute, Korea Advanced Institute of Science and Technology
• Original Assignee: Korea Advanced Institute of Science and Technology
• Inventors: Kim, Hyoung Gon, Kim, Ig Jae, Chang, Joon-Hyuk, Bae, Chang Seok, Lee, Kye Hwan
• Primary Examiner(s): SAINT CYR, LEONARD
• Application Number:
  US12/711,030
• Publication Number:
  US 20100217595A1
• Time in Patent Office: 812 Days
• Field of Search: None
• US Class Current: 704/250
• CPC Class Codes: G10L 17/26 Recognition of special voice characteristics, e.g. for use in lie detectors; Recognition of animal voices

Abstract

Disclosed herein is a method for emotion recognition based on a minimum classification error. In the method, a speaker'"'"'s neutral emotion is extracted using a Gaussian mixture model (GMM), other emotions except the neutral emotion are classified using the Gaussian Mixture Model to which a discriminative weight for minimizing the loss function of a classification error for the feature vector for emotion recognition is applied. In the emotion recognition, the emotion recognition is performed by applying a discriminative weight evaluated using the Gaussian Mixture Model based on minimum classification error to feature vectors of the emotion classified with difficult, thereby enhancing the performance of emotion recognition.

Claims

1. A method for emotion recognition based on a minimum classification error, the method comprising:

extracting a feature vector for emotion recognition based on a voice signal generated from a speaker and a galvanic skin response of the speaker, the feature vector for emotion recognition including a voice signal feature vector containing information extracted from the voice signal of the speaker and a galvanic skin response feature vector extracted from galvanic skin response of the speaker;

classifying a neutral emotion using a Gaussian mixture model based on the extracted feature vector for emotion recognition; and

classifying other emotions except the previously classified neutral emotion using the Gaussian Mixture Model to which a discriminative weight for minimizing the loss function of a classification error for the feature vector for emotion recognition is applied, wherein the emotions are classified by comparing a likelihood ratio with a threshold value, and the likelihood ratio is obtained from the Gaussian Mixture Model modified by the discriminative weight.

2. The method according to claim 1, wherein the voice signal feature vector includes at least one or more feature vectors of mel frequency cepstral coefficient (MFCC), delta MFCC (Δ

MFCC), pitch, music continuity counter (MCC).

3. The method according to claim 1, wherein the Gaussian Mixture Model is generated by extracting one or more feature vectors of MFCC, delta MFCC, pitch and MCC and a feature vector of galvanic skin responses, which are feature vectors for training, from database comprising voice signals and galvanic skin responses.

4. The method according to claim 1, wherein other emotions except the neutral emotion are angry and joy.

5. The method according to claim 1, wherein classifying other emotions except the previously classified neutral emotion comprises:

obtaining optimum discriminative weight for minimizing the loss function based on training data; and

applying the optimum discriminative weight to the Gaussian Mixture Model for emotion recognition.

6. The method according to claim 5, further comprising:

classifying emotions between angry and joy using the Gaussian Mixture Model modified by the optimum discriminative weight.