A Speaker Dependent Emotion Recognition Framework

1
A Speaker Dependent Emotion Recognition Framework

• T. P. Kostoulas
• N. Fakotakis
• Artificial Intelligence Group
• Wire Communications Laboratory
• Dep. Electrical Computer Engineering
• University of Patras
• July 2006

2
Outline
Problem Definition
State of the art
System Description
Experimental Setup - Results
Conclusions
3
Problem Definition

• Individuals, as emotional beings, tend to
  interact with other emotional beings. Therefore
  information about the end-users emotional state
  is necessary.
• Affective Computing
• User friendly interfaces
• Affective interaction with dialogue systems

4
Affective Computing
5
Emotion Recognitions Factors

• Automatic Emotion Recognitions factors
• Feature Space
• Classifier
• Speech Corpus
• Emotion Categories

6
State of the art (1)
7
State of the art (2)
8
System Description
9
Database Description

• Database used LDC-2002S28 Emotional Prosody
  Database
• Recordings Dates and numbers uttered in the
  appropriate emotional category
• Eight Actors
• Five Females
• Three Males
• Fifteen Emotions

10
Experimental Setup (1)

• Five Basic Emotions chosen
• Hot Anger
• Sad
• Happy
• Panic
• Neutral

• Leave one out technique had been used

11
Experimental Setup (2)

• Data Labeling
• Default label defined through the creation of the
  database.
• Human Evaluation
• 4 Greek listeners, 2 males and 2 females with
  proficient English level.

12
Experimental Results (1)
Mean Accuracy 85.40
13
Experimental Results (2)
14
Experimental Results (3)
15
Experimental Results (4)
Mean Accuracy 78.39
16
Conclusions
Real world-based Experimental Setup is
challenging
Acted Speech was used
No phoneme recognizer or linguistic information
Human Computer 7.01
Neutral Anger 98.13
Panic Anger 39.99
5 Emotions 85.40
17
Thank You !
tkost_at_wcl.ee.upatras.gr fakotaki_at_wcl.ee.upatras.gr
http//www.wcl.ee.upatras.gr/ai