Speech emotion recognition using machine learning techniques: Feature extraction and comparison of convolutional...

Area of Science:

• Computational Linguistics
• Machine Learning
• Signal Processing

Background:

• Speech is a primary channel for conveying emotions.
• Accurate speech emotion recognition (SER) is crucial for human-computer interaction.
• Limited datasets pose challenges for developing robust SER models.

Purpose of the Study:

• To classify emotions in speech using diverse audio features and machine learning models.
• To evaluate the performance of Random Forest (RF) with feature selection against a one-dimensional Convolutional Neural Network (conv1D).
• To investigate common misclassifications in SER.

Main Methods:

• Extraction of audio features: Mel-frequency cepstral coefficients, chromogram, Mel-scale spectrogram, spectral contrast, Tonnetz, and zero-crossing rate.
• Dataset augmentation and combination of all audio files prior to analysis.
• Comparison of RF with feature selection and conv1D models for SER.

Main Results:

• RF with feature selection achieved a higher average accuracy (69%) compared to conv1D.
• RF demonstrated the highest precision for fear (72%) and recall for calm (84%).
• Common misclassifications included anger with happy, disgust with sad/neutral, and fear with sad, attributed to acoustic feature similarities.

Conclusions:

• RF with feature selection is effective for SER, especially with limited datasets.
• The study highlights the potential of acoustic feature analysis for emotion classification.
• Understanding misclassification patterns can guide future improvements in SER systems.