Time-frequency feature representation using multi-resolution texture analysis and acoustic activity detector for...

Area of Science:

• Speech processing
• Human-Computer Interaction (HCI)
• Affective computing

Background:

• Emotional speech classification is crucial for human-computer interaction.
• Existing methods often lack accuracy in real-world, naturally occurring emotional speech.
• Emotions manifest differently across various frequency bands, necessitating nuanced feature extraction.

Purpose of the Study:

• To introduce a novel feature extraction technique called Multi-Resolution Texture Image Information (MRTII).
• To enhance the characterization and classification of emotions in speech signals using multi-resolution texture analysis.
• To improve the accuracy of emotional speech recognition, particularly in real-life applications.

Main Methods:

• Feature extraction using Multi-Resolution Texture Image Information (MRTII) derived from speech spectrograms.
• Application of an Acoustic Activity Detection (AAD) algorithm for enhanced discrimination.
• Utilizing two corpora of naturally-occurring dialogs from real-life call centers.
• Comparison with traditional Mel-Frequency Cepstral Coefficients (MFCC) and state-of-the-art features.

Main Results:

• MRTII features demonstrate superior performance compared to MFCC and other state-of-the-art features.
• The proposed method achieves improved correct classification rates across different language databases.
• Experimental results validate the effectiveness of MRTII for real-life emotional speech recognition.

Conclusions:

• MRTII, inspired by human visual perception of spectrograms, offers significant advantages for emotional speech classification.
• The multi-resolution texture analysis provides clearer discrimination between emotions than uniform-resolution methods.
• The approach shows promise for robust and accurate real-life emotional recognition systems.