A multi-dilated convolution network for speech emotion recognition.

Area of Science:

• Artificial Intelligence
• Affective Computing
• Speech Processing

Background:

• Speech emotion recognition (SER) is crucial for human-computer interaction.
• Deep Neural Networks (DNNs) and Convolutional Neural Networks (CNNs) show promise for SER using speech spectrograms.
• Existing methods face challenges in effectively learning deep patterns from spectrograms.

Purpose of the Study:

• To propose a novel SER model leveraging utterance-level spectrogram analysis.
• To enhance feature extraction by integrating Spatial Pyramid Pooling (SPP) and attention mechanisms.
• To adapt techniques from face recognition, specifically ArcFace, for improved SER performance.

Main Methods:

• Utilized Spatial Pyramid Pooling (SPP) to overcome CNN size constraints.
• Extracted global and multi-local level feature vectors using SPP.
• Implemented an attention model to weigh extracted feature vectors.
• Applied the ArcFace layer, typically used in face recognition, to the SER task.

Main Results:

• Achieved an unweighted accuracy of 67.9% on the IEMOCAP dataset.
• Achieved an unweighted accuracy of 77.6% on the EMODB dataset.
• Demonstrated improved SER performance through the proposed model architecture.

Conclusions:

• The novel SER model effectively learns deep patterns from utterance-level spectrograms.
• Integrating SPP, attention, and ArcFace enhances SER accuracy.
• The proposed approach offers a promising direction for advancing SER technology.