🎙️ Speech Emotion Recognition using IEMOCAP & CREMA-D — EmoNet Model

Overview

This repository contains the implementation and experimental work corresponding to the conference paper:
“Speech Emotion Recognition using Dilated CNN-BiLSTM Network” (ECCE 2025) by Md. Nur Alam.

The project focuses on building a robust speech emotion recognition (SER) system using IEMOCAP and CREMA-D datasets.
It includes preprocessing, data balancing, augmentation, baseline modeling, and the development of the EmoNet model — a hybrid deep learning architecture combining Dilated CNN and BiLSTM layers to improve temporal and spectral emotion recognition.

---

Key Features

• Multi-dataset training using IEMOCAP and CREMA-D
• Feature extraction: MFCC and Log-Mel spectrograms
• Data preprocessing: cleaning, balancing, and augmentation
• Baseline model: EmoTech (CNN-BiLSTM hybrid)
• Proposed model: EmoNet — combines Dilated CNN + BiLSTM for capturing short-term and long-range dependencies
• Implemented in Python using TensorFlow/Keras

---

Repository Structure

📁 Speech_Emotion_Recognition/ ├── ECCE_2025_Speech_Emotion_Recognition_Nur.pdf # Conference paper ├── Paper_IEMOCAP.ipynb # Main notebook with code ├── README.md # This file

---

Methodology

1. Data Preprocessing

• Combined IEMOCAP and CREMA-D datasets
• Extracted MFCC and Log-Mel features
• Applied data balancing to handle class imbalance
• Performed augmentation including time-shifting, pitch scaling, and noise addition

2. Baseline Model — EmoTech

A CNN-BiLSTM hybrid model serving as the initial benchmark. Focuses on sequential temporal learning but lacks global context.

3. Proposed Model — EmoNet

A Dilated CNN + BiLSTM network designed to capture both short-term and long-range dependencies in emotional speech.

• Dilated convolutions expand context without losing resolution
• BiLSTM layers model bidirectional temporal relationships
• Dropout and BatchNorm layers improve stability and generalization

---

---

Author

Md. Nur Alam
Department of Electronics and Telecommunication Engineering, CUET, Bangladesh
📧 nuralam1908046@gmail.com
🔗 LinkedIn

Citation

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

@inproceedings{alam2025emotech,
  author    = {N. Alam and N. Mobassara and N. Mamun},
  title     = {Revamping EmoTech: Leveraging Dilated Convolution for Cutting-Edge Robust Speech Emotion Recognition},
  booktitle = {2025 International Conference on Electrical, Computer and Communication Engineering (ECCE)},
  year      = {2025},
  pages     = {1--6},
  address   = {Chittagong, Bangladesh},
  doi       = {10.1109/ECCE64574.2025.11013956},
  keywords  = {Emotion recognition, Accuracy, Convolution, Bidirectional long short term memory, Speech recognition, Data augmentation, Feature extraction, Data models, Robustness, Mel frequency cepstral coefficient, Speech Emotion Recognition, Dilated Convolution, BiLSTM, MFCC}
}