On the Speech Properties and Feature Extraction Methods in Speech Emotion Recognition

Area of Science:

• Speech processing and recognition
• Computational linguistics
• Affective computing

Background:

• Existing speech emotion recognition (SER) systems lack detailed understanding of how specific acoustic features and processing choices influence performance.
• A need exists to analyze fundamental speech characteristics and modeling techniques for improved SER system design.

Purpose of the Study:

• To extend the physical perspective on SER by analyzing basic speech characteristics and modeling methods.
• To investigate the impact of various signal processing techniques on SER accuracy.
• To identify optimal settings for robust emotion recognition from speech.

Main Methods:

• Analysis of time characteristics (segmentation, windowing), frequency ranges/scales, spectrograms, vocal tract modeling (filter banks, LPC), and excitation signals.
• Evaluation of cepstral features, magnitude/phase manipulations, and advanced classification methods.
• Application of rigorous statistical tests including N-fold cross-validation, paired t-tests, rank, and Pearson correlations.

Main Results:

• Several settings achieved approximately 75% accuracy in recognizing seven distinct emotions.
• Vocal tract features utilizing psychoacoustic filter banks (0-8 kHz) yielded the most successful results.
• Spectrograms containing vocal tract and excitation information also performed well.
• Basic processing steps like pre-emphasis and segmentation significantly influenced outcomes.

Conclusions:

• The study provides crucial insights into the physical underpinnings of SER, guiding future system development.
• Optimal performance is linked to specific vocal tract feature extraction and processing parameters.
• Even elementary signal processing choices have a substantial effect on SER accuracy, with findings demonstrating robustness across datasets.