MelTrans: Mel-Spectrogram Relationship-Learning for Speech Emotion Recognition via Transformers.

Area of Science:

• Artificial Intelligence
• Human-Computer Interaction
• Signal Processing

Background:

• Speech emotion recognition (SER) is crucial for natural human-computer interaction.
• Existing SER methods struggle with subtle emotions and noisy environments.
• Advanced feature extraction and dependency modeling are needed for robust SER.

Purpose of the Study:

• To introduce MelTrans, a novel Transformer-based model for enhanced speech emotion recognition.
• To address challenges in detecting subtle emotional nuances and recognizing emotions in noisy speech.
• To improve the accuracy and robustness of SER systems.

Main Methods:

• Developed MelTrans, a dual-stream Transformer-based model.
• Utilized speech mel-spectrograms to capture broad dependencies.
• Focused on learning core features and long-range dependencies within speech data.

Main Results:

• MelTrans achieved 92.52% accuracy on the EmoDB dataset.
• MelTrans achieved 76.54% accuracy on the IEMOCAP dataset.
• Demonstrated superior performance in capturing critical cues and long-range dependencies.

Conclusions:

• MelTrans effectively addresses complex challenges in speech emotion recognition.
• The model sets new benchmarks for SER on the EmoDB and IEMOCAP datasets.
• Highlights the potential of Transformer architectures for nuanced emotion detection in speech.