Effect on speech emotion classification of a feature selection approach using a convolutional neural network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Speech emotion recognition (SER) is complex due to the difficulty in identifying effective classification features.
• Traditional handcrafted features often prove insufficient for accurate speaker emotion identification.
• Deep learning models offer potential advantages for feature extraction in SER.

Purpose of the Study:

• To investigate the effectiveness of deep convolutional neural networks (DCNNs) for speech emotion recognition.
• To apply a feature selection (FS) approach to identify discriminative features for SER.
• To evaluate the performance of various classifiers using selected features across multiple speech emotion databases.

Main Methods:

• Utilized a pretrained deep convolutional neural network (DCNN) framework for feature extraction from speech emotion datasets.
• Implemented a feature selection (FS) method to pinpoint the most discriminative features for emotion classification.
• Employed multiple classification algorithms including Random Forest (RF), Decision Tree (DT), Support Vector Machine (SVM), Multilayer Perceptron (MLP), and K-Nearest Neighbors (KNN) to classify seven emotions.
• Conducted experiments on four publicly accessible databases: Emo-DB, SAVEE, RAVDESS, and IEMOCAP.

Main Results:

• Achieved high speaker-dependent (SD) recognition accuracies: 92.02% (Emo-DB), 88.77% (SAVEE), 93.61% (RAVDESS), and 77.23% (IEMOCAP) with the FS method.
• Demonstrated superior performance for speaker-independent (SI) SER compared to existing handcrafted feature-based methods.
• Observed that all classifiers on EMO-DB achieved over 80% accuracy, irrespective of the feature selection technique.

Conclusions:

• The proposed DCNN-based feature extraction combined with feature selection offers a robust approach for speech emotion recognition.
• The method significantly enhances classification accuracy, particularly for speaker-independent scenarios.
• The findings highlight the importance of discriminative feature identification for advancing SER technology.