Speech Emotion Recognition via CNN-Transformer and Multidimensional Attention Mechanism

📖 Introduction

This repository implements a Speech Emotion Recognition (SER) system using a hybrid ​​CNN-Transformer architecture​​ enhanced with a ​​Multidimensional Attention Mechanism​​.

🧠 Model Architecture

1. Overall Framework

🛠️ Installation

Clone the repository and create the conda environment:

git clone https://github.com/SCNU-RISLAB/CTMAM.git
cd CTMAM_SER_master

conda create -n CTMAM python=3.10
conda activate CTMAM
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=10.2 -c pytorch
pip install -r requirements.txt

📂 Data Process

IEMOCAP Dataset

python process_IEMPCAP.py

python process_Emodb.py

🚀 Training

IEMOCAP Dataset

python train_IEMOCAP.py

python process_Emodb.py

📜 Citation

If you use this work in your research, please cite:

bibtex

@article{tang2025speech,

title = {Speech emotion recognition via cnn-transformer and multidimensional attention mechanism},

author = {Tang, Xiaoyu and Huang, Jiazheng and Lin, Yixin and Dang, Ting and Cheng, Jintao},

journal = {Speech Communication},

pages = {103242},

year = {2025},

publisher = {Elsevier}

}