A New Network Structure for Speech Emotion Recognition Research.

Area of Science:

• Artificial Intelligence
• Speech Processing
• Machine Learning

Background:

• Speech emotion recognition (SER) is crucial for human-computer interaction.
• Extracting relevant acoustic features is key to SER performance.
• Current methods struggle to integrate dispersed local and global emotional information in speech signals.

Purpose of the Study:

• To develop a novel network model for comprehensive acoustic feature extraction in SER.
• To address the challenge of integrating local and global emotional information in speech.
• To enhance the accuracy and generalization of speech emotion recognition and sentiment analysis.

Main Methods:

• Proposed a deep learning model integrating a gated recurrent unit (GRU) and multi-head attention.
• Utilized a bidirectional GRU (Bi-GRU) to capture temporal dependencies.
• Employed multi-head attention to focus on salient acoustic features.
• Evaluated the model on IEMOCAP and Emo-DB speech emotion corpora.
• Tested generalization on CH-SIMS and MOSI speech sentiment analysis datasets.

Main Results:

• The Bi-GRU and multi-head attention model significantly outperformed traditional methods on multiple evaluation metrics for SER.
• The model demonstrated excellent generalization capabilities in speech sentiment analysis tasks.
• Achieved superior performance in detecting emotional nuances within speech signals.

Conclusions:

• The proposed Bi-GRU and multi-head attention model effectively integrates local and global acoustic information for improved SER.
• This approach offers a significant advancement in speech emotion recognition and sentiment analysis.
• The model's strong performance highlights its potential for real-world applications in understanding emotional content in speech.