Year: 2023
Type: article
Abstract: With technological advancements, machine needs to understand human speech, i.e., Human-Computer Interaction (HCI) has become vital. For natural interaction, emotion detection in speech is a must. Time domain features can identify a few emotions, whereas some are determined by inherently using frequency domain features. With wavelet-based features majority of emotion discriminating features can be dentified. A common observation is that happy emotion is seen to be majorly misclassified as an angry emotion. Reduction in this misclassification is achieved with the proposed feature vector. Spectral features and Discrete Wavelet Transform (DWT) features form the proposed feature vector. Feature selection is made using statistical test analysis of variance ANOVA. The model is verified using SVM and MLP classifiers. In this work, a speech emotion recognition system is evaluated using a German audio database (EMODB). It is seen to be able to recognize happy and angry emotions with better accuracy as compared to state-of-the-art algorithms. For four emotion classes: happy, angry, neutral and sad proposed model performance with DWT features has improved by 3% compared to baseline features in the case of both the classifiers, viz. SVM and MLP
Authors: Ashwini S. Shinde, Vaishali V. Patil
Language: English
Topic: Emotion and Mood Recognition
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4388754381