Pre-trained Deep Convolution Neural Network Model With Attention for Speech Emotion Recognition.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Speech emotion recognition (SER) is challenging due to speaker variability.
• Effective feature extraction and classification models are crucial for SER performance.

Purpose of the Study:

• To propose a novel DCNN-BLSTMwA model for improved speech emotion recognition.
• To enhance feature extraction and classification for accurate emotion detection from speech.

Main Methods:

• Data preprocessing included enhancement and balancing.
• Log Mel-spectrograms (static, delta, delta-delta) were used as input for DCNN.
• A DCNN pre-trained on ImageNet generated segment-level features, stacked into utterance-level features.
• Bidirectional Long Short-Term Memory (BLSTM) with an attention layer learned high-level emotional features.
• A Deep Neural Network (DNN) predicted the final emotion.

Main Results:

• The DCNN-BLSTMwA model achieved an unweighted average recall (UAR) of 87.86% on the EMO-DB dataset.
• The model obtained a UAR of 68.50% on the IEMOCAP dataset.
• Performance surpassed most existing SER methods, demonstrating the model's effectiveness.

Conclusions:

• The proposed DCNN-BLSTMwA method is effective for speech emotion recognition.
• The integration of DCNN, BLSTM with attention, and DNN significantly improves SER accuracy.
• This approach offers a robust solution for recognizing emotions in speech signals.