Representation learning with parameterised quantum circuits for advancing speech emotion recognition.

Area of Science:

• Quantum Computing
• Machine Learning
• Signal Processing
• Affective Computing

Background:

• Speech emotion recognition (SER) is challenging due to subtle vocal variations.
• Representation learning in complex signals requires advanced techniques.
• Quantum machine learning (QML) shows potential for improving signal processing tasks.

Purpose of the Study:

• Investigate the use of parameterised quantum circuits (PQCs) for SER.
• Develop a hybrid quantum-classical architecture for emotion recognition.
• Evaluate the effectiveness of QML in enhancing emotional feature representations.

Main Methods:

• Proposed a hybrid model integrating PQCs with a convolutional neural network (CNN).
• Leveraged quantum properties like superposition and entanglement for feature enrichment.
• Conducted experiments on benchmark datasets: IEMOCAP, RECOLA, and MSP-IMPROV.

Main Results:

• The hybrid quantum-classical model outperformed a purely classical CNN baseline in classification performance.
• Achieved over 50% reduction in trainable parameters compared to the classical CNN.
• Adjusted Rand Index (ARI) analysis showed improved alignment of quantum-derived features with true emotion classes.

Conclusions:

• QML demonstrates potential for enhancing emotion recognition systems.
• The hybrid approach offers improved performance and efficiency for SER.
• This study lays groundwork for future quantum-enhanced affective computing systems.