Emotional Speech Recognition Using Deep Neural Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Speech Processing

Background:

• Human emotional expression is crucial for communication, with vocal cues being a primary focus of study.
• Recognizing emotions in speech is vital for human-computer interaction and affective computing.

Purpose of the Study:

• To investigate the effectiveness of deep neural networks (CNN, CRNN, GRU) for speech emotion recognition.
• To evaluate model performance on the Interactive Emotional Dyadic Motion Capture (IEMOCAP) corpus.

Main Methods:

• Utilized Mel spectral coefficients and speech signal parameters for feature extraction.
• Employed data augmentation techniques, including voice modification and white noise addition.
• Trained and compared Convolutional Neural Networks (CNN), Convolutional Recurrent Neural Networks (CRNN), and Gated Recurrent Units (GRU) models.

Main Results:

• The Gated Recurrent Unit (GRU) model achieved the highest average recognition accuracy of 97.47%.
• GRU model performance surpassed existing studies on speech emotion recognition using the IEMOCAP corpus.

Conclusions:

• Deep neural networks, particularly GRU, show significant promise for accurate speech emotion recognition.
• The proposed method offers a superior approach for identifying emotions in speech data.