Research on Chinese Speech Emotion Recognition Based on Deep Neural Network and Acoustic Features

Area of Science:

• Artificial Intelligence
• Speech Processing
• Natural Language Processing

Background:

• Emotion recognition is gaining traction in AI applications.
• Voice-controlled interfaces are becoming prevalent in smart homes and AI robots.
• Current systems often rely on touch-sensitive interfaces, limiting natural interaction.

Purpose of the Study:

• To develop a Chinese speech emotion recognition system.
• To enable smart home voice assistants and AI service robots to understand user emotions via voice.
• To transition from touch-based to voice-based interaction for AI systems.

Main Methods:

• Developed a specialized Deep Neural Network (DNN) model.
• Utilized 29 acoustic features from acoustic theory for training.
• Implemented audio enhancement techniques: waveform adjustment, pitch adjustment, and pre-emphasis to augment datasets and improve accuracy.

Main Results:

• Achieved an average emotion recognition accuracy of 88.9% on the CASIA Chinese sentiment corpus.
• Demonstrated the effectiveness of the proposed deep learning model and audio adjustment methods.
• Successfully identified emotions in Chinese short sentences.

Conclusions:

• The proposed deep learning model and audio enhancement techniques are effective for Chinese speech emotion recognition.
• The system can be integrated into Chinese voice assistants and dialogue applications.
• This research facilitates more natural and intuitive human-AI interaction through voice emotion understanding.